blob: 248dd119a86e2ce643ea4e33e63fccee4d41dd32 [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>
Rusty Russell9adef582007-05-08 00:26:42 -070058#include <linux/module.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>
62
Jakub Jelinek4732efb2005-09-06 15:16:25 -070063#include <asm/futex.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070064
Ingo Molnarc87e2832006-06-27 02:54:58 -070065#include "rtmutex_common.h"
66
Thomas Gleixnera0c1e902008-02-23 15:23:57 -080067int __read_mostly futex_cmpxchg_enabled;
68
Linus Torvalds1da177e2005-04-16 15:20:36 -070069#define FUTEX_HASHBITS (CONFIG_BASE_SMALL ? 4 : 8)
70
71/*
Ingo Molnarc87e2832006-06-27 02:54:58 -070072 * Priority Inheritance state:
73 */
74struct futex_pi_state {
75 /*
76 * list of 'owned' pi_state instances - these have to be
77 * cleaned up in do_exit() if the task exits prematurely:
78 */
79 struct list_head list;
80
81 /*
82 * The PI object:
83 */
84 struct rt_mutex pi_mutex;
85
86 struct task_struct *owner;
87 atomic_t refcount;
88
89 union futex_key key;
90};
91
92/*
Linus Torvalds1da177e2005-04-16 15:20:36 -070093 * We use this hashed waitqueue instead of a normal wait_queue_t, so
94 * we can wake only the relevant ones (hashed queues may be shared).
95 *
96 * A futex_q has a woken state, just like tasks have TASK_RUNNING.
Pierre Peifferec92d082007-05-09 02:35:00 -070097 * It is considered woken when plist_node_empty(&q->list) || q->lock_ptr == 0.
Linus Torvalds1da177e2005-04-16 15:20:36 -070098 * The order of wakup is always to make the first condition true, then
Darren Hart73500ac2008-12-17 17:29:56 -080099 * wake up q->waiter, then make the second condition true.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700100 */
101struct futex_q {
Pierre Peifferec92d082007-05-09 02:35:00 -0700102 struct plist_node list;
Thomas Gleixnerf1a11e02009-05-05 19:21:40 +0200103 /* Waiter reference */
104 struct task_struct *task;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700105
Ingo Molnare2970f22006-06-27 02:54:47 -0700106 /* Which hash list lock to use: */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700107 spinlock_t *lock_ptr;
108
Ingo Molnare2970f22006-06-27 02:54:47 -0700109 /* Key which the futex is hashed on: */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700110 union futex_key key;
111
Ingo Molnarc87e2832006-06-27 02:54:58 -0700112 /* Optional priority inheritance state: */
113 struct futex_pi_state *pi_state;
Thomas Gleixnercd689982008-02-01 17:45:14 +0100114
Darren Hart52400ba2009-04-03 13:40:49 -0700115 /* rt_waiter storage for requeue_pi: */
116 struct rt_mutex_waiter *rt_waiter;
117
Darren Hart84bc4af2009-08-13 17:36:53 -0700118 /* The expected requeue pi target futex key: */
119 union futex_key *requeue_pi_key;
120
Thomas Gleixnercd689982008-02-01 17:45:14 +0100121 /* Bitset for the optional bitmasked wakeup */
122 u32 bitset;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700123};
124
125/*
Darren Hartb2d09942009-03-12 00:55:37 -0700126 * Hash buckets are shared by all the futex_keys that hash to the same
127 * location. Each key may have multiple futex_q structures, one for each task
128 * waiting on a futex.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700129 */
130struct futex_hash_bucket {
Pierre Peifferec92d082007-05-09 02:35:00 -0700131 spinlock_t lock;
132 struct plist_head chain;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700133};
134
135static struct futex_hash_bucket futex_queues[1<<FUTEX_HASHBITS];
136
Linus Torvalds1da177e2005-04-16 15:20:36 -0700137/*
138 * We hash on the keys returned from get_futex_key (see below).
139 */
140static struct futex_hash_bucket *hash_futex(union futex_key *key)
141{
142 u32 hash = jhash2((u32*)&key->both.word,
143 (sizeof(key->both.word)+sizeof(key->both.ptr))/4,
144 key->both.offset);
145 return &futex_queues[hash & ((1 << FUTEX_HASHBITS)-1)];
146}
147
148/*
149 * Return 1 if two futex_keys are equal, 0 otherwise.
150 */
151static inline int match_futex(union futex_key *key1, union futex_key *key2)
152{
153 return (key1->both.word == key2->both.word
154 && key1->both.ptr == key2->both.ptr
155 && key1->both.offset == key2->both.offset);
156}
157
Peter Zijlstra38d47c12008-09-26 19:32:20 +0200158/*
159 * Take a reference to the resource addressed by a key.
160 * Can be called while holding spinlocks.
161 *
162 */
163static void get_futex_key_refs(union futex_key *key)
164{
165 if (!key->both.ptr)
166 return;
167
168 switch (key->both.offset & (FUT_OFF_INODE|FUT_OFF_MMSHARED)) {
169 case FUT_OFF_INODE:
170 atomic_inc(&key->shared.inode->i_count);
171 break;
172 case FUT_OFF_MMSHARED:
173 atomic_inc(&key->private.mm->mm_count);
174 break;
175 }
176}
177
178/*
179 * Drop a reference to the resource addressed by a key.
180 * The hash bucket spinlock must not be held.
181 */
182static void drop_futex_key_refs(union futex_key *key)
183{
Darren Hart90621c42008-12-29 19:43:21 -0800184 if (!key->both.ptr) {
185 /* If we're here then we tried to put a key we failed to get */
186 WARN_ON_ONCE(1);
Peter Zijlstra38d47c12008-09-26 19:32:20 +0200187 return;
Darren Hart90621c42008-12-29 19:43:21 -0800188 }
Peter Zijlstra38d47c12008-09-26 19:32:20 +0200189
190 switch (key->both.offset & (FUT_OFF_INODE|FUT_OFF_MMSHARED)) {
191 case FUT_OFF_INODE:
192 iput(key->shared.inode);
193 break;
194 case FUT_OFF_MMSHARED:
195 mmdrop(key->private.mm);
196 break;
197 }
198}
199
Eric Dumazet34f01cc2007-05-09 02:35:04 -0700200/**
201 * get_futex_key - Get parameters which are the keys for a futex.
202 * @uaddr: virtual address of the futex
Darren Hartb2d09942009-03-12 00:55:37 -0700203 * @fshared: 0 for a PROCESS_PRIVATE futex, 1 for PROCESS_SHARED
Eric Dumazet34f01cc2007-05-09 02:35:04 -0700204 * @key: address where result is stored.
Thomas Gleixner64d13042009-05-18 21:20:10 +0200205 * @rw: mapping needs to be read/write (values: VERIFY_READ, VERIFY_WRITE)
Eric Dumazet34f01cc2007-05-09 02:35:04 -0700206 *
207 * Returns a negative error code or 0
208 * The key words are stored in *key on success.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700209 *
Josef "Jeff" Sipekf3a43f32006-12-08 02:36:43 -0800210 * For shared mappings, it's (page->index, vma->vm_file->f_path.dentry->d_inode,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700211 * offset_within_page). For private mappings, it's (uaddr, current->mm).
212 * We can usually work out the index without swapping in the page.
213 *
Darren Hartb2d09942009-03-12 00:55:37 -0700214 * lock_page() might sleep, the caller should not hold a spinlock.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700215 */
Thomas Gleixner64d13042009-05-18 21:20:10 +0200216static int
217get_futex_key(u32 __user *uaddr, int fshared, union futex_key *key, int rw)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700218{
Ingo Molnare2970f22006-06-27 02:54:47 -0700219 unsigned long address = (unsigned long)uaddr;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700220 struct mm_struct *mm = current->mm;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700221 struct page *page;
222 int err;
223
224 /*
225 * The futex address must be "naturally" aligned.
226 */
Ingo Molnare2970f22006-06-27 02:54:47 -0700227 key->both.offset = address % PAGE_SIZE;
Eric Dumazet34f01cc2007-05-09 02:35:04 -0700228 if (unlikely((address % sizeof(u32)) != 0))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700229 return -EINVAL;
Ingo Molnare2970f22006-06-27 02:54:47 -0700230 address -= key->both.offset;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700231
232 /*
Eric Dumazet34f01cc2007-05-09 02:35:04 -0700233 * PROCESS_PRIVATE futexes are fast.
234 * As the mm cannot disappear under us and the 'key' only needs
235 * virtual address, we dont even have to find the underlying vma.
236 * Note : We do have to check 'uaddr' is a valid user address,
237 * but access_ok() should be faster than find_vma()
238 */
239 if (!fshared) {
Thomas Gleixner64d13042009-05-18 21:20:10 +0200240 if (unlikely(!access_ok(rw, uaddr, sizeof(u32))))
Eric Dumazet34f01cc2007-05-09 02:35:04 -0700241 return -EFAULT;
242 key->private.mm = mm;
243 key->private.address = address;
Peter Zijlstra42569c32008-09-30 12:33:07 +0200244 get_futex_key_refs(key);
Eric Dumazet34f01cc2007-05-09 02:35:04 -0700245 return 0;
246 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700247
Peter Zijlstra38d47c12008-09-26 19:32:20 +0200248again:
Thomas Gleixner64d13042009-05-18 21:20:10 +0200249 err = get_user_pages_fast(address, 1, rw == VERIFY_WRITE, &page);
Peter Zijlstra38d47c12008-09-26 19:32:20 +0200250 if (err < 0)
251 return err;
252
Sonny Raoce2ae532009-07-10 18:13:13 -0500253 page = compound_head(page);
Peter Zijlstra38d47c12008-09-26 19:32:20 +0200254 lock_page(page);
255 if (!page->mapping) {
256 unlock_page(page);
257 put_page(page);
258 goto again;
259 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700260
261 /*
262 * Private mappings are handled in a simple way.
263 *
264 * NOTE: When userspace waits on a MAP_SHARED mapping, even if
265 * it's a read-only handle, it's expected that futexes attach to
Peter Zijlstra38d47c12008-09-26 19:32:20 +0200266 * the object not the particular process.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700267 */
Peter Zijlstra38d47c12008-09-26 19:32:20 +0200268 if (PageAnon(page)) {
269 key->both.offset |= FUT_OFF_MMSHARED; /* ref taken on mm */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700270 key->private.mm = mm;
Ingo Molnare2970f22006-06-27 02:54:47 -0700271 key->private.address = address;
Peter Zijlstra38d47c12008-09-26 19:32:20 +0200272 } else {
273 key->both.offset |= FUT_OFF_INODE; /* inode-based key */
274 key->shared.inode = page->mapping->host;
275 key->shared.pgoff = page->index;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700276 }
277
Peter Zijlstra38d47c12008-09-26 19:32:20 +0200278 get_futex_key_refs(key);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700279
Peter Zijlstra38d47c12008-09-26 19:32:20 +0200280 unlock_page(page);
281 put_page(page);
282 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700283}
284
Peter Zijlstra38d47c12008-09-26 19:32:20 +0200285static inline
Peter Zijlstrac2f9f202008-09-26 19:32:23 +0200286void put_futex_key(int fshared, union futex_key *key)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700287{
Peter Zijlstra38d47c12008-09-26 19:32:20 +0200288 drop_futex_key_refs(key);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700289}
290
Thomas Gleixnerd0725992009-06-11 23:15:43 +0200291/*
292 * fault_in_user_writeable - fault in user address and verify RW access
293 * @uaddr: pointer to faulting user space address
294 *
295 * Slow path to fixup the fault we just took in the atomic write
296 * access to @uaddr.
297 *
298 * We have no generic implementation of a non destructive write to the
299 * user address. We know that we faulted in the atomic pagefault
300 * disabled section so we can as well avoid the #PF overhead by
301 * calling get_user_pages() right away.
302 */
303static int fault_in_user_writeable(u32 __user *uaddr)
304{
305 int ret = get_user_pages(current, current->mm, (unsigned long)uaddr,
Thomas Gleixneraa715282009-06-25 14:27:58 +0200306 1, 1, 0, NULL, NULL);
Thomas Gleixnerd0725992009-06-11 23:15:43 +0200307 return ret < 0 ? ret : 0;
308}
309
Darren Hart4b1c4862009-04-03 13:39:42 -0700310/**
311 * futex_top_waiter() - Return the highest priority waiter on a futex
312 * @hb: the hash bucket the futex_q's reside in
313 * @key: the futex key (to distinguish it from other futex futex_q's)
314 *
315 * Must be called with the hb lock held.
316 */
317static struct futex_q *futex_top_waiter(struct futex_hash_bucket *hb,
318 union futex_key *key)
319{
320 struct futex_q *this;
321
322 plist_for_each_entry(this, &hb->chain, list) {
323 if (match_futex(&this->key, key))
324 return this;
325 }
326 return NULL;
327}
328
Thomas Gleixner36cf3b52007-07-15 23:41:20 -0700329static u32 cmpxchg_futex_value_locked(u32 __user *uaddr, u32 uval, u32 newval)
330{
331 u32 curval;
332
333 pagefault_disable();
334 curval = futex_atomic_cmpxchg_inatomic(uaddr, uval, newval);
335 pagefault_enable();
336
337 return curval;
338}
339
340static int get_futex_value_locked(u32 *dest, u32 __user *from)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700341{
342 int ret;
343
Peter Zijlstraa8663742006-12-06 20:32:20 -0800344 pagefault_disable();
Ingo Molnare2970f22006-06-27 02:54:47 -0700345 ret = __copy_from_user_inatomic(dest, from, sizeof(u32));
Peter Zijlstraa8663742006-12-06 20:32:20 -0800346 pagefault_enable();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700347
348 return ret ? -EFAULT : 0;
349}
350
Ingo Molnarc87e2832006-06-27 02:54:58 -0700351
352/*
353 * PI code:
354 */
355static int refill_pi_state_cache(void)
356{
357 struct futex_pi_state *pi_state;
358
359 if (likely(current->pi_state_cache))
360 return 0;
361
Burman Yan4668edc2006-12-06 20:38:51 -0800362 pi_state = kzalloc(sizeof(*pi_state), GFP_KERNEL);
Ingo Molnarc87e2832006-06-27 02:54:58 -0700363
364 if (!pi_state)
365 return -ENOMEM;
366
Ingo Molnarc87e2832006-06-27 02:54:58 -0700367 INIT_LIST_HEAD(&pi_state->list);
368 /* pi_mutex gets initialized later */
369 pi_state->owner = NULL;
370 atomic_set(&pi_state->refcount, 1);
Peter Zijlstra38d47c12008-09-26 19:32:20 +0200371 pi_state->key = FUTEX_KEY_INIT;
Ingo Molnarc87e2832006-06-27 02:54:58 -0700372
373 current->pi_state_cache = pi_state;
374
375 return 0;
376}
377
378static struct futex_pi_state * alloc_pi_state(void)
379{
380 struct futex_pi_state *pi_state = current->pi_state_cache;
381
382 WARN_ON(!pi_state);
383 current->pi_state_cache = NULL;
384
385 return pi_state;
386}
387
388static void free_pi_state(struct futex_pi_state *pi_state)
389{
390 if (!atomic_dec_and_test(&pi_state->refcount))
391 return;
392
393 /*
394 * If pi_state->owner is NULL, the owner is most probably dying
395 * and has cleaned up the pi_state already
396 */
397 if (pi_state->owner) {
398 spin_lock_irq(&pi_state->owner->pi_lock);
399 list_del_init(&pi_state->list);
400 spin_unlock_irq(&pi_state->owner->pi_lock);
401
402 rt_mutex_proxy_unlock(&pi_state->pi_mutex, pi_state->owner);
403 }
404
405 if (current->pi_state_cache)
406 kfree(pi_state);
407 else {
408 /*
409 * pi_state->list is already empty.
410 * clear pi_state->owner.
411 * refcount is at 0 - put it back to 1.
412 */
413 pi_state->owner = NULL;
414 atomic_set(&pi_state->refcount, 1);
415 current->pi_state_cache = pi_state;
416 }
417}
418
419/*
420 * Look up the task based on what TID userspace gave us.
421 * We dont trust it.
422 */
423static struct task_struct * futex_find_get_task(pid_t pid)
424{
425 struct task_struct *p;
David Howellsc69e8d92008-11-14 10:39:19 +1100426 const struct cred *cred = current_cred(), *pcred;
Ingo Molnarc87e2832006-06-27 02:54:58 -0700427
Oleg Nesterovd359b542006-09-29 02:00:55 -0700428 rcu_read_lock();
Pavel Emelyanov228ebcb2007-10-18 23:40:16 -0700429 p = find_task_by_vpid(pid);
David Howellsc69e8d92008-11-14 10:39:19 +1100430 if (!p) {
Thomas Gleixnera06381f2007-06-23 11:48:40 +0200431 p = ERR_PTR(-ESRCH);
David Howellsc69e8d92008-11-14 10:39:19 +1100432 } else {
433 pcred = __task_cred(p);
434 if (cred->euid != pcred->euid &&
435 cred->euid != pcred->uid)
436 p = ERR_PTR(-ESRCH);
437 else
438 get_task_struct(p);
439 }
Thomas Gleixnera06381f2007-06-23 11:48:40 +0200440
Oleg Nesterovd359b542006-09-29 02:00:55 -0700441 rcu_read_unlock();
Ingo Molnarc87e2832006-06-27 02:54:58 -0700442
443 return p;
444}
445
446/*
447 * This task is holding PI mutexes at exit time => bad.
448 * Kernel cleans up PI-state, but userspace is likely hosed.
449 * (Robust-futex cleanup is separate and might save the day for userspace.)
450 */
451void exit_pi_state_list(struct task_struct *curr)
452{
Ingo Molnarc87e2832006-06-27 02:54:58 -0700453 struct list_head *next, *head = &curr->pi_state_list;
454 struct futex_pi_state *pi_state;
Ingo Molnar627371d2006-07-29 05:16:20 +0200455 struct futex_hash_bucket *hb;
Peter Zijlstra38d47c12008-09-26 19:32:20 +0200456 union futex_key key = FUTEX_KEY_INIT;
Ingo Molnarc87e2832006-06-27 02:54:58 -0700457
Thomas Gleixnera0c1e902008-02-23 15:23:57 -0800458 if (!futex_cmpxchg_enabled)
459 return;
Ingo Molnarc87e2832006-06-27 02:54:58 -0700460 /*
461 * We are a ZOMBIE and nobody can enqueue itself on
462 * pi_state_list anymore, but we have to be careful
Ingo Molnar627371d2006-07-29 05:16:20 +0200463 * versus waiters unqueueing themselves:
Ingo Molnarc87e2832006-06-27 02:54:58 -0700464 */
465 spin_lock_irq(&curr->pi_lock);
466 while (!list_empty(head)) {
467
468 next = head->next;
469 pi_state = list_entry(next, struct futex_pi_state, list);
470 key = pi_state->key;
Ingo Molnar627371d2006-07-29 05:16:20 +0200471 hb = hash_futex(&key);
Ingo Molnarc87e2832006-06-27 02:54:58 -0700472 spin_unlock_irq(&curr->pi_lock);
473
Ingo Molnarc87e2832006-06-27 02:54:58 -0700474 spin_lock(&hb->lock);
475
476 spin_lock_irq(&curr->pi_lock);
Ingo Molnar627371d2006-07-29 05:16:20 +0200477 /*
478 * We dropped the pi-lock, so re-check whether this
479 * task still owns the PI-state:
480 */
Ingo Molnarc87e2832006-06-27 02:54:58 -0700481 if (head->next != next) {
482 spin_unlock(&hb->lock);
483 continue;
484 }
485
Ingo Molnarc87e2832006-06-27 02:54:58 -0700486 WARN_ON(pi_state->owner != curr);
Ingo Molnar627371d2006-07-29 05:16:20 +0200487 WARN_ON(list_empty(&pi_state->list));
488 list_del_init(&pi_state->list);
Ingo Molnarc87e2832006-06-27 02:54:58 -0700489 pi_state->owner = NULL;
490 spin_unlock_irq(&curr->pi_lock);
491
492 rt_mutex_unlock(&pi_state->pi_mutex);
493
494 spin_unlock(&hb->lock);
495
496 spin_lock_irq(&curr->pi_lock);
497 }
498 spin_unlock_irq(&curr->pi_lock);
499}
500
501static int
Pierre Peifferd0aa7a72007-05-09 02:35:02 -0700502lookup_pi_state(u32 uval, struct futex_hash_bucket *hb,
503 union futex_key *key, struct futex_pi_state **ps)
Ingo Molnarc87e2832006-06-27 02:54:58 -0700504{
505 struct futex_pi_state *pi_state = NULL;
506 struct futex_q *this, *next;
Pierre Peifferec92d082007-05-09 02:35:00 -0700507 struct plist_head *head;
Ingo Molnarc87e2832006-06-27 02:54:58 -0700508 struct task_struct *p;
Alexey Kuznetsov778e9a92007-06-08 13:47:00 -0700509 pid_t pid = uval & FUTEX_TID_MASK;
Ingo Molnarc87e2832006-06-27 02:54:58 -0700510
511 head = &hb->chain;
512
Pierre Peifferec92d082007-05-09 02:35:00 -0700513 plist_for_each_entry_safe(this, next, head, list) {
Pierre Peifferd0aa7a72007-05-09 02:35:02 -0700514 if (match_futex(&this->key, key)) {
Ingo Molnarc87e2832006-06-27 02:54:58 -0700515 /*
516 * Another waiter already exists - bump up
517 * the refcount and return its pi_state:
518 */
519 pi_state = this->pi_state;
Thomas Gleixner06a9ec22006-07-10 04:44:30 -0700520 /*
521 * Userspace might have messed up non PI and PI futexes
522 */
523 if (unlikely(!pi_state))
524 return -EINVAL;
525
Ingo Molnar627371d2006-07-29 05:16:20 +0200526 WARN_ON(!atomic_read(&pi_state->refcount));
Alexey Kuznetsov778e9a92007-06-08 13:47:00 -0700527 WARN_ON(pid && pi_state->owner &&
528 pi_state->owner->pid != pid);
Ingo Molnar627371d2006-07-29 05:16:20 +0200529
Ingo Molnarc87e2832006-06-27 02:54:58 -0700530 atomic_inc(&pi_state->refcount);
Pierre Peifferd0aa7a72007-05-09 02:35:02 -0700531 *ps = pi_state;
Ingo Molnarc87e2832006-06-27 02:54:58 -0700532
533 return 0;
534 }
535 }
536
537 /*
Ingo Molnare3f2dde2006-07-29 05:17:57 +0200538 * We are the first waiter - try to look up the real owner and attach
Alexey Kuznetsov778e9a92007-06-08 13:47:00 -0700539 * the new pi_state to it, but bail out when TID = 0
Ingo Molnarc87e2832006-06-27 02:54:58 -0700540 */
Alexey Kuznetsov778e9a92007-06-08 13:47:00 -0700541 if (!pid)
Ingo Molnare3f2dde2006-07-29 05:17:57 +0200542 return -ESRCH;
Ingo Molnarc87e2832006-06-27 02:54:58 -0700543 p = futex_find_get_task(pid);
Alexey Kuznetsov778e9a92007-06-08 13:47:00 -0700544 if (IS_ERR(p))
545 return PTR_ERR(p);
546
547 /*
548 * We need to look at the task state flags to figure out,
549 * whether the task is exiting. To protect against the do_exit
550 * change of the task flags, we do this protected by
551 * p->pi_lock:
552 */
553 spin_lock_irq(&p->pi_lock);
554 if (unlikely(p->flags & PF_EXITING)) {
555 /*
556 * The task is on the way out. When PF_EXITPIDONE is
557 * set, we know that the task has finished the
558 * cleanup:
559 */
560 int ret = (p->flags & PF_EXITPIDONE) ? -ESRCH : -EAGAIN;
561
562 spin_unlock_irq(&p->pi_lock);
563 put_task_struct(p);
564 return ret;
565 }
Ingo Molnarc87e2832006-06-27 02:54:58 -0700566
567 pi_state = alloc_pi_state();
568
569 /*
570 * Initialize the pi_mutex in locked state and make 'p'
571 * the owner of it:
572 */
573 rt_mutex_init_proxy_locked(&pi_state->pi_mutex, p);
574
575 /* Store the key for possible exit cleanups: */
Pierre Peifferd0aa7a72007-05-09 02:35:02 -0700576 pi_state->key = *key;
Ingo Molnarc87e2832006-06-27 02:54:58 -0700577
Ingo Molnar627371d2006-07-29 05:16:20 +0200578 WARN_ON(!list_empty(&pi_state->list));
Ingo Molnarc87e2832006-06-27 02:54:58 -0700579 list_add(&pi_state->list, &p->pi_state_list);
580 pi_state->owner = p;
581 spin_unlock_irq(&p->pi_lock);
582
583 put_task_struct(p);
584
Pierre Peifferd0aa7a72007-05-09 02:35:02 -0700585 *ps = pi_state;
Ingo Molnarc87e2832006-06-27 02:54:58 -0700586
587 return 0;
588}
589
Darren Hart1a520842009-04-03 13:39:52 -0700590/**
591 * futex_lock_pi_atomic() - atomic work required to acquire a pi aware futex
Darren Hartbab5bc92009-04-07 23:23:50 -0700592 * @uaddr: the pi futex user address
593 * @hb: the pi futex hash bucket
594 * @key: the futex key associated with uaddr and hb
595 * @ps: the pi_state pointer where we store the result of the
596 * lookup
597 * @task: the task to perform the atomic lock work for. This will
598 * be "current" except in the case of requeue pi.
599 * @set_waiters: force setting the FUTEX_WAITERS bit (1) or not (0)
Darren Hart1a520842009-04-03 13:39:52 -0700600 *
601 * Returns:
602 * 0 - ready to wait
603 * 1 - acquired the lock
604 * <0 - error
605 *
606 * The hb->lock and futex_key refs shall be held by the caller.
607 */
608static int futex_lock_pi_atomic(u32 __user *uaddr, struct futex_hash_bucket *hb,
609 union futex_key *key,
610 struct futex_pi_state **ps,
Darren Hartbab5bc92009-04-07 23:23:50 -0700611 struct task_struct *task, int set_waiters)
Darren Hart1a520842009-04-03 13:39:52 -0700612{
613 int lock_taken, ret, ownerdied = 0;
614 u32 uval, newval, curval;
615
616retry:
617 ret = lock_taken = 0;
618
619 /*
620 * To avoid races, we attempt to take the lock here again
621 * (by doing a 0 -> TID atomic cmpxchg), while holding all
622 * the locks. It will most likely not succeed.
623 */
624 newval = task_pid_vnr(task);
Darren Hartbab5bc92009-04-07 23:23:50 -0700625 if (set_waiters)
626 newval |= FUTEX_WAITERS;
Darren Hart1a520842009-04-03 13:39:52 -0700627
628 curval = cmpxchg_futex_value_locked(uaddr, 0, newval);
629
630 if (unlikely(curval == -EFAULT))
631 return -EFAULT;
632
633 /*
634 * Detect deadlocks.
635 */
636 if ((unlikely((curval & FUTEX_TID_MASK) == task_pid_vnr(task))))
637 return -EDEADLK;
638
639 /*
640 * Surprise - we got the lock. Just return to userspace:
641 */
642 if (unlikely(!curval))
643 return 1;
644
645 uval = curval;
646
647 /*
648 * Set the FUTEX_WAITERS flag, so the owner will know it has someone
649 * to wake at the next unlock.
650 */
651 newval = curval | FUTEX_WAITERS;
652
653 /*
654 * There are two cases, where a futex might have no owner (the
655 * owner TID is 0): OWNER_DIED. We take over the futex in this
656 * case. We also do an unconditional take over, when the owner
657 * of the futex died.
658 *
659 * This is safe as we are protected by the hash bucket lock !
660 */
661 if (unlikely(ownerdied || !(curval & FUTEX_TID_MASK))) {
662 /* Keep the OWNER_DIED bit */
663 newval = (curval & ~FUTEX_TID_MASK) | task_pid_vnr(task);
664 ownerdied = 0;
665 lock_taken = 1;
666 }
667
668 curval = cmpxchg_futex_value_locked(uaddr, uval, newval);
669
670 if (unlikely(curval == -EFAULT))
671 return -EFAULT;
672 if (unlikely(curval != uval))
673 goto retry;
674
675 /*
676 * We took the lock due to owner died take over.
677 */
678 if (unlikely(lock_taken))
679 return 1;
680
681 /*
682 * We dont have the lock. Look up the PI state (or create it if
683 * we are the first waiter):
684 */
685 ret = lookup_pi_state(uval, hb, key, ps);
686
687 if (unlikely(ret)) {
688 switch (ret) {
689 case -ESRCH:
690 /*
691 * No owner found for this futex. Check if the
692 * OWNER_DIED bit is set to figure out whether
693 * this is a robust futex or not.
694 */
695 if (get_futex_value_locked(&curval, uaddr))
696 return -EFAULT;
697
698 /*
699 * We simply start over in case of a robust
700 * futex. The code above will take the futex
701 * and return happy.
702 */
703 if (curval & FUTEX_OWNER_DIED) {
704 ownerdied = 1;
705 goto retry;
706 }
707 default:
708 break;
709 }
710 }
711
712 return ret;
713}
714
Ingo Molnarc87e2832006-06-27 02:54:58 -0700715/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700716 * The hash bucket lock must be held when this is called.
717 * Afterwards, the futex_q must not be accessed.
718 */
719static void wake_futex(struct futex_q *q)
720{
Thomas Gleixnerf1a11e02009-05-05 19:21:40 +0200721 struct task_struct *p = q->task;
722
723 /*
724 * We set q->lock_ptr = NULL _before_ we wake up the task. If
725 * a non futex wake up happens on another CPU then the task
726 * might exit and p would dereference a non existing task
727 * struct. Prevent this by holding a reference on p across the
728 * wake up.
729 */
730 get_task_struct(p);
731
Pierre Peifferec92d082007-05-09 02:35:00 -0700732 plist_del(&q->list, &q->list.plist);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700733 /*
Thomas Gleixnerf1a11e02009-05-05 19:21:40 +0200734 * The waiting task can free the futex_q as soon as
735 * q->lock_ptr = NULL is written, without taking any locks. A
736 * memory barrier is required here to prevent the following
737 * store to lock_ptr from getting ahead of the plist_del.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700738 */
Ralf Baechleccdea2f2006-12-06 20:40:26 -0800739 smp_wmb();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700740 q->lock_ptr = NULL;
Thomas Gleixnerf1a11e02009-05-05 19:21:40 +0200741
742 wake_up_state(p, TASK_NORMAL);
743 put_task_struct(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700744}
745
Ingo Molnarc87e2832006-06-27 02:54:58 -0700746static int wake_futex_pi(u32 __user *uaddr, u32 uval, struct futex_q *this)
747{
748 struct task_struct *new_owner;
749 struct futex_pi_state *pi_state = this->pi_state;
750 u32 curval, newval;
751
752 if (!pi_state)
753 return -EINVAL;
754
Ingo Molnar217788672007-03-16 13:38:31 -0800755 spin_lock(&pi_state->pi_mutex.wait_lock);
Ingo Molnarc87e2832006-06-27 02:54:58 -0700756 new_owner = rt_mutex_next_owner(&pi_state->pi_mutex);
757
758 /*
759 * This happens when we have stolen the lock and the original
760 * pending owner did not enqueue itself back on the rt_mutex.
761 * Thats not a tragedy. We know that way, that a lock waiter
762 * is on the fly. We make the futex_q waiter the pending owner.
763 */
764 if (!new_owner)
765 new_owner = this->task;
766
767 /*
768 * We pass it to the next owner. (The WAITERS bit is always
769 * kept enabled while there is PI state around. We must also
770 * preserve the owner died bit.)
771 */
Ingo Molnare3f2dde2006-07-29 05:17:57 +0200772 if (!(uval & FUTEX_OWNER_DIED)) {
Alexey Kuznetsov778e9a92007-06-08 13:47:00 -0700773 int ret = 0;
774
Pavel Emelyanovb4888932007-10-18 23:40:14 -0700775 newval = FUTEX_WAITERS | task_pid_vnr(new_owner);
Ingo Molnarc87e2832006-06-27 02:54:58 -0700776
Thomas Gleixner36cf3b52007-07-15 23:41:20 -0700777 curval = cmpxchg_futex_value_locked(uaddr, uval, newval);
Alexey Kuznetsov778e9a92007-06-08 13:47:00 -0700778
Ingo Molnare3f2dde2006-07-29 05:17:57 +0200779 if (curval == -EFAULT)
Alexey Kuznetsov778e9a92007-06-08 13:47:00 -0700780 ret = -EFAULT;
Thomas Gleixnercde898f2007-12-05 15:46:09 +0100781 else if (curval != uval)
Alexey Kuznetsov778e9a92007-06-08 13:47:00 -0700782 ret = -EINVAL;
783 if (ret) {
784 spin_unlock(&pi_state->pi_mutex.wait_lock);
785 return ret;
786 }
Ingo Molnare3f2dde2006-07-29 05:17:57 +0200787 }
Ingo Molnarc87e2832006-06-27 02:54:58 -0700788
Ingo Molnar627371d2006-07-29 05:16:20 +0200789 spin_lock_irq(&pi_state->owner->pi_lock);
790 WARN_ON(list_empty(&pi_state->list));
791 list_del_init(&pi_state->list);
792 spin_unlock_irq(&pi_state->owner->pi_lock);
793
794 spin_lock_irq(&new_owner->pi_lock);
795 WARN_ON(!list_empty(&pi_state->list));
Ingo Molnarc87e2832006-06-27 02:54:58 -0700796 list_add(&pi_state->list, &new_owner->pi_state_list);
797 pi_state->owner = new_owner;
Ingo Molnar627371d2006-07-29 05:16:20 +0200798 spin_unlock_irq(&new_owner->pi_lock);
799
Ingo Molnar217788672007-03-16 13:38:31 -0800800 spin_unlock(&pi_state->pi_mutex.wait_lock);
Ingo Molnarc87e2832006-06-27 02:54:58 -0700801 rt_mutex_unlock(&pi_state->pi_mutex);
802
803 return 0;
804}
805
806static int unlock_futex_pi(u32 __user *uaddr, u32 uval)
807{
808 u32 oldval;
809
810 /*
811 * There is no waiter, so we unlock the futex. The owner died
812 * bit has not to be preserved here. We are the owner:
813 */
Thomas Gleixner36cf3b52007-07-15 23:41:20 -0700814 oldval = cmpxchg_futex_value_locked(uaddr, uval, 0);
Ingo Molnarc87e2832006-06-27 02:54:58 -0700815
816 if (oldval == -EFAULT)
817 return oldval;
818 if (oldval != uval)
819 return -EAGAIN;
820
821 return 0;
822}
823
Linus Torvalds1da177e2005-04-16 15:20:36 -0700824/*
Ingo Molnar8b8f3192006-07-03 00:25:05 -0700825 * Express the locking dependencies for lockdep:
826 */
827static inline void
828double_lock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
829{
830 if (hb1 <= hb2) {
831 spin_lock(&hb1->lock);
832 if (hb1 < hb2)
833 spin_lock_nested(&hb2->lock, SINGLE_DEPTH_NESTING);
834 } else { /* hb1 > hb2 */
835 spin_lock(&hb2->lock);
836 spin_lock_nested(&hb1->lock, SINGLE_DEPTH_NESTING);
837 }
838}
839
Darren Hart5eb3dc62009-03-12 00:55:52 -0700840static inline void
841double_unlock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
842{
Darren Hartf061d352009-03-12 15:11:18 -0700843 spin_unlock(&hb1->lock);
Ingo Molnar88f502f2009-03-13 10:32:07 +0100844 if (hb1 != hb2)
845 spin_unlock(&hb2->lock);
Darren Hart5eb3dc62009-03-12 00:55:52 -0700846}
847
Ingo Molnar8b8f3192006-07-03 00:25:05 -0700848/*
Darren Hartb2d09942009-03-12 00:55:37 -0700849 * Wake up waiters matching bitset queued on this futex (uaddr).
Linus Torvalds1da177e2005-04-16 15:20:36 -0700850 */
Peter Zijlstrac2f9f202008-09-26 19:32:23 +0200851static int futex_wake(u32 __user *uaddr, int fshared, int nr_wake, u32 bitset)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700852{
Ingo Molnare2970f22006-06-27 02:54:47 -0700853 struct futex_hash_bucket *hb;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700854 struct futex_q *this, *next;
Pierre Peifferec92d082007-05-09 02:35:00 -0700855 struct plist_head *head;
Peter Zijlstra38d47c12008-09-26 19:32:20 +0200856 union futex_key key = FUTEX_KEY_INIT;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700857 int ret;
858
Thomas Gleixnercd689982008-02-01 17:45:14 +0100859 if (!bitset)
860 return -EINVAL;
861
Thomas Gleixner64d13042009-05-18 21:20:10 +0200862 ret = get_futex_key(uaddr, fshared, &key, VERIFY_READ);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700863 if (unlikely(ret != 0))
864 goto out;
865
Ingo Molnare2970f22006-06-27 02:54:47 -0700866 hb = hash_futex(&key);
867 spin_lock(&hb->lock);
868 head = &hb->chain;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700869
Pierre Peifferec92d082007-05-09 02:35:00 -0700870 plist_for_each_entry_safe(this, next, head, list) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700871 if (match_futex (&this->key, &key)) {
Darren Hart52400ba2009-04-03 13:40:49 -0700872 if (this->pi_state || this->rt_waiter) {
Ingo Molnared6f7b12006-07-01 04:35:46 -0700873 ret = -EINVAL;
874 break;
875 }
Thomas Gleixnercd689982008-02-01 17:45:14 +0100876
877 /* Check if one of the bits is set in both bitsets */
878 if (!(this->bitset & bitset))
879 continue;
880
Linus Torvalds1da177e2005-04-16 15:20:36 -0700881 wake_futex(this);
882 if (++ret >= nr_wake)
883 break;
884 }
885 }
886
Ingo Molnare2970f22006-06-27 02:54:47 -0700887 spin_unlock(&hb->lock);
Peter Zijlstra38d47c12008-09-26 19:32:20 +0200888 put_futex_key(fshared, &key);
Darren Hart42d35d42008-12-29 15:49:53 -0800889out:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700890 return ret;
891}
892
893/*
Jakub Jelinek4732efb2005-09-06 15:16:25 -0700894 * Wake up all waiters hashed on the physical page that is mapped
895 * to this virtual address:
896 */
Ingo Molnare2970f22006-06-27 02:54:47 -0700897static int
Peter Zijlstrac2f9f202008-09-26 19:32:23 +0200898futex_wake_op(u32 __user *uaddr1, int fshared, u32 __user *uaddr2,
Ingo Molnare2970f22006-06-27 02:54:47 -0700899 int nr_wake, int nr_wake2, int op)
Jakub Jelinek4732efb2005-09-06 15:16:25 -0700900{
Peter Zijlstra38d47c12008-09-26 19:32:20 +0200901 union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
Ingo Molnare2970f22006-06-27 02:54:47 -0700902 struct futex_hash_bucket *hb1, *hb2;
Pierre Peifferec92d082007-05-09 02:35:00 -0700903 struct plist_head *head;
Jakub Jelinek4732efb2005-09-06 15:16:25 -0700904 struct futex_q *this, *next;
Darren Harte4dc5b72009-03-12 00:56:13 -0700905 int ret, op_ret;
Jakub Jelinek4732efb2005-09-06 15:16:25 -0700906
Darren Harte4dc5b72009-03-12 00:56:13 -0700907retry:
Thomas Gleixner64d13042009-05-18 21:20:10 +0200908 ret = get_futex_key(uaddr1, fshared, &key1, VERIFY_READ);
Jakub Jelinek4732efb2005-09-06 15:16:25 -0700909 if (unlikely(ret != 0))
910 goto out;
Thomas Gleixner64d13042009-05-18 21:20:10 +0200911 ret = get_futex_key(uaddr2, fshared, &key2, VERIFY_WRITE);
Jakub Jelinek4732efb2005-09-06 15:16:25 -0700912 if (unlikely(ret != 0))
Darren Hart42d35d42008-12-29 15:49:53 -0800913 goto out_put_key1;
Jakub Jelinek4732efb2005-09-06 15:16:25 -0700914
Ingo Molnare2970f22006-06-27 02:54:47 -0700915 hb1 = hash_futex(&key1);
916 hb2 = hash_futex(&key2);
Jakub Jelinek4732efb2005-09-06 15:16:25 -0700917
Ingo Molnar8b8f3192006-07-03 00:25:05 -0700918 double_lock_hb(hb1, hb2);
Darren Harte4dc5b72009-03-12 00:56:13 -0700919retry_private:
Ingo Molnare2970f22006-06-27 02:54:47 -0700920 op_ret = futex_atomic_op_inuser(op, uaddr2);
Jakub Jelinek4732efb2005-09-06 15:16:25 -0700921 if (unlikely(op_ret < 0)) {
Jakub Jelinek4732efb2005-09-06 15:16:25 -0700922
Darren Hart5eb3dc62009-03-12 00:55:52 -0700923 double_unlock_hb(hb1, hb2);
Jakub Jelinek4732efb2005-09-06 15:16:25 -0700924
David Howells7ee1dd32006-01-06 00:11:44 -0800925#ifndef CONFIG_MMU
Ingo Molnare2970f22006-06-27 02:54:47 -0700926 /*
927 * we don't get EFAULT from MMU faults if we don't have an MMU,
928 * but we might get them from range checking
929 */
David Howells7ee1dd32006-01-06 00:11:44 -0800930 ret = op_ret;
Darren Hart42d35d42008-12-29 15:49:53 -0800931 goto out_put_keys;
David Howells7ee1dd32006-01-06 00:11:44 -0800932#endif
933
David Gibson796f8d92005-11-07 00:59:33 -0800934 if (unlikely(op_ret != -EFAULT)) {
935 ret = op_ret;
Darren Hart42d35d42008-12-29 15:49:53 -0800936 goto out_put_keys;
David Gibson796f8d92005-11-07 00:59:33 -0800937 }
938
Thomas Gleixnerd0725992009-06-11 23:15:43 +0200939 ret = fault_in_user_writeable(uaddr2);
Jakub Jelinek4732efb2005-09-06 15:16:25 -0700940 if (ret)
Darren Hartde87fcc2009-03-12 00:55:46 -0700941 goto out_put_keys;
Jakub Jelinek4732efb2005-09-06 15:16:25 -0700942
Darren Harte4dc5b72009-03-12 00:56:13 -0700943 if (!fshared)
944 goto retry_private;
945
Darren Hartde87fcc2009-03-12 00:55:46 -0700946 put_futex_key(fshared, &key2);
947 put_futex_key(fshared, &key1);
Darren Harte4dc5b72009-03-12 00:56:13 -0700948 goto retry;
Jakub Jelinek4732efb2005-09-06 15:16:25 -0700949 }
950
Ingo Molnare2970f22006-06-27 02:54:47 -0700951 head = &hb1->chain;
Jakub Jelinek4732efb2005-09-06 15:16:25 -0700952
Pierre Peifferec92d082007-05-09 02:35:00 -0700953 plist_for_each_entry_safe(this, next, head, list) {
Jakub Jelinek4732efb2005-09-06 15:16:25 -0700954 if (match_futex (&this->key, &key1)) {
955 wake_futex(this);
956 if (++ret >= nr_wake)
957 break;
958 }
959 }
960
961 if (op_ret > 0) {
Ingo Molnare2970f22006-06-27 02:54:47 -0700962 head = &hb2->chain;
Jakub Jelinek4732efb2005-09-06 15:16:25 -0700963
964 op_ret = 0;
Pierre Peifferec92d082007-05-09 02:35:00 -0700965 plist_for_each_entry_safe(this, next, head, list) {
Jakub Jelinek4732efb2005-09-06 15:16:25 -0700966 if (match_futex (&this->key, &key2)) {
967 wake_futex(this);
968 if (++op_ret >= nr_wake2)
969 break;
970 }
971 }
972 ret += op_ret;
973 }
974
Darren Hart5eb3dc62009-03-12 00:55:52 -0700975 double_unlock_hb(hb1, hb2);
Darren Hart42d35d42008-12-29 15:49:53 -0800976out_put_keys:
Peter Zijlstra38d47c12008-09-26 19:32:20 +0200977 put_futex_key(fshared, &key2);
Darren Hart42d35d42008-12-29 15:49:53 -0800978out_put_key1:
Peter Zijlstra38d47c12008-09-26 19:32:20 +0200979 put_futex_key(fshared, &key1);
Darren Hart42d35d42008-12-29 15:49:53 -0800980out:
Jakub Jelinek4732efb2005-09-06 15:16:25 -0700981 return ret;
982}
983
Darren Hart9121e472009-04-03 13:40:31 -0700984/**
985 * requeue_futex() - Requeue a futex_q from one hb to another
986 * @q: the futex_q to requeue
987 * @hb1: the source hash_bucket
988 * @hb2: the target hash_bucket
989 * @key2: the new key for the requeued futex_q
990 */
991static inline
992void requeue_futex(struct futex_q *q, struct futex_hash_bucket *hb1,
993 struct futex_hash_bucket *hb2, union futex_key *key2)
994{
995
996 /*
997 * If key1 and key2 hash to the same bucket, no need to
998 * requeue.
999 */
1000 if (likely(&hb1->chain != &hb2->chain)) {
1001 plist_del(&q->list, &hb1->chain);
1002 plist_add(&q->list, &hb2->chain);
1003 q->lock_ptr = &hb2->lock;
1004#ifdef CONFIG_DEBUG_PI_LIST
1005 q->list.plist.lock = &hb2->lock;
1006#endif
1007 }
1008 get_futex_key_refs(key2);
1009 q->key = *key2;
1010}
1011
Darren Hart52400ba2009-04-03 13:40:49 -07001012/**
1013 * requeue_pi_wake_futex() - Wake a task that acquired the lock during requeue
1014 * q: the futex_q
1015 * key: the key of the requeue target futex
Darren Hartbeda2c72009-08-09 15:34:39 -07001016 * hb: the hash_bucket of the requeue target futex
Darren Hart52400ba2009-04-03 13:40:49 -07001017 *
1018 * During futex_requeue, with requeue_pi=1, it is possible to acquire the
1019 * target futex if it is uncontended or via a lock steal. Set the futex_q key
1020 * to the requeue target futex so the waiter can detect the wakeup on the right
1021 * futex, but remove it from the hb and NULL the rt_waiter so it can detect
Darren Hartbeda2c72009-08-09 15:34:39 -07001022 * atomic lock acquisition. Set the q->lock_ptr to the requeue target hb->lock
1023 * to protect access to the pi_state to fixup the owner later. Must be called
1024 * with both q->lock_ptr and hb->lock held.
Darren Hart52400ba2009-04-03 13:40:49 -07001025 */
1026static inline
Darren Hartbeda2c72009-08-09 15:34:39 -07001027void requeue_pi_wake_futex(struct futex_q *q, union futex_key *key,
1028 struct futex_hash_bucket *hb)
Darren Hart52400ba2009-04-03 13:40:49 -07001029{
1030 drop_futex_key_refs(&q->key);
1031 get_futex_key_refs(key);
1032 q->key = *key;
1033
1034 WARN_ON(plist_node_empty(&q->list));
1035 plist_del(&q->list, &q->list.plist);
1036
1037 WARN_ON(!q->rt_waiter);
1038 q->rt_waiter = NULL;
1039
Darren Hartbeda2c72009-08-09 15:34:39 -07001040 q->lock_ptr = &hb->lock;
1041#ifdef CONFIG_DEBUG_PI_LIST
1042 q->list.plist.lock = &hb->lock;
1043#endif
1044
Thomas Gleixnerf1a11e02009-05-05 19:21:40 +02001045 wake_up_state(q->task, TASK_NORMAL);
Darren Hart52400ba2009-04-03 13:40:49 -07001046}
1047
1048/**
1049 * futex_proxy_trylock_atomic() - Attempt an atomic lock for the top waiter
Darren Hartbab5bc92009-04-07 23:23:50 -07001050 * @pifutex: the user address of the to futex
1051 * @hb1: the from futex hash bucket, must be locked by the caller
1052 * @hb2: the to futex hash bucket, must be locked by the caller
1053 * @key1: the from futex key
1054 * @key2: the to futex key
1055 * @ps: address to store the pi_state pointer
1056 * @set_waiters: force setting the FUTEX_WAITERS bit (1) or not (0)
Darren Hart52400ba2009-04-03 13:40:49 -07001057 *
1058 * Try and get the lock on behalf of the top waiter if we can do it atomically.
Darren Hartbab5bc92009-04-07 23:23:50 -07001059 * Wake the top waiter if we succeed. If the caller specified set_waiters,
1060 * then direct futex_lock_pi_atomic() to force setting the FUTEX_WAITERS bit.
1061 * hb1 and hb2 must be held by the caller.
Darren Hart52400ba2009-04-03 13:40:49 -07001062 *
1063 * Returns:
1064 * 0 - failed to acquire the lock atomicly
1065 * 1 - acquired the lock
1066 * <0 - error
1067 */
1068static int futex_proxy_trylock_atomic(u32 __user *pifutex,
1069 struct futex_hash_bucket *hb1,
1070 struct futex_hash_bucket *hb2,
1071 union futex_key *key1, union futex_key *key2,
Darren Hartbab5bc92009-04-07 23:23:50 -07001072 struct futex_pi_state **ps, int set_waiters)
Darren Hart52400ba2009-04-03 13:40:49 -07001073{
Darren Hartbab5bc92009-04-07 23:23:50 -07001074 struct futex_q *top_waiter = NULL;
Darren Hart52400ba2009-04-03 13:40:49 -07001075 u32 curval;
1076 int ret;
1077
1078 if (get_futex_value_locked(&curval, pifutex))
1079 return -EFAULT;
1080
Darren Hartbab5bc92009-04-07 23:23:50 -07001081 /*
1082 * Find the top_waiter and determine if there are additional waiters.
1083 * If the caller intends to requeue more than 1 waiter to pifutex,
1084 * force futex_lock_pi_atomic() to set the FUTEX_WAITERS bit now,
1085 * as we have means to handle the possible fault. If not, don't set
1086 * the bit unecessarily as it will force the subsequent unlock to enter
1087 * the kernel.
1088 */
Darren Hart52400ba2009-04-03 13:40:49 -07001089 top_waiter = futex_top_waiter(hb1, key1);
1090
1091 /* There are no waiters, nothing for us to do. */
1092 if (!top_waiter)
1093 return 0;
1094
Darren Hart84bc4af2009-08-13 17:36:53 -07001095 /* Ensure we requeue to the expected futex. */
1096 if (!match_futex(top_waiter->requeue_pi_key, key2))
1097 return -EINVAL;
1098
Darren Hart52400ba2009-04-03 13:40:49 -07001099 /*
Darren Hartbab5bc92009-04-07 23:23:50 -07001100 * Try to take the lock for top_waiter. Set the FUTEX_WAITERS bit in
1101 * the contended case or if set_waiters is 1. The pi_state is returned
1102 * in ps in contended cases.
Darren Hart52400ba2009-04-03 13:40:49 -07001103 */
Darren Hartbab5bc92009-04-07 23:23:50 -07001104 ret = futex_lock_pi_atomic(pifutex, hb2, key2, ps, top_waiter->task,
1105 set_waiters);
Darren Hart52400ba2009-04-03 13:40:49 -07001106 if (ret == 1)
Darren Hartbeda2c72009-08-09 15:34:39 -07001107 requeue_pi_wake_futex(top_waiter, key2, hb2);
Darren Hart52400ba2009-04-03 13:40:49 -07001108
1109 return ret;
1110}
1111
1112/**
1113 * futex_requeue() - Requeue waiters from uaddr1 to uaddr2
1114 * uaddr1: source futex user address
1115 * uaddr2: target futex user address
1116 * nr_wake: number of waiters to wake (must be 1 for requeue_pi)
1117 * nr_requeue: number of waiters to requeue (0-INT_MAX)
1118 * requeue_pi: if we are attempting to requeue from a non-pi futex to a
1119 * pi futex (pi to pi requeue is not supported)
1120 *
1121 * Requeue waiters on uaddr1 to uaddr2. In the requeue_pi case, try to acquire
1122 * uaddr2 atomically on behalf of the top waiter.
1123 *
1124 * Returns:
1125 * >=0 - on success, the number of tasks requeued or woken
1126 * <0 - on error
Linus Torvalds1da177e2005-04-16 15:20:36 -07001127 */
Peter Zijlstrac2f9f202008-09-26 19:32:23 +02001128static int futex_requeue(u32 __user *uaddr1, int fshared, u32 __user *uaddr2,
Darren Hart52400ba2009-04-03 13:40:49 -07001129 int nr_wake, int nr_requeue, u32 *cmpval,
1130 int requeue_pi)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001131{
Peter Zijlstra38d47c12008-09-26 19:32:20 +02001132 union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
Darren Hart52400ba2009-04-03 13:40:49 -07001133 int drop_count = 0, task_count = 0, ret;
1134 struct futex_pi_state *pi_state = NULL;
Ingo Molnare2970f22006-06-27 02:54:47 -07001135 struct futex_hash_bucket *hb1, *hb2;
Pierre Peifferec92d082007-05-09 02:35:00 -07001136 struct plist_head *head1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001137 struct futex_q *this, *next;
Darren Hart52400ba2009-04-03 13:40:49 -07001138 u32 curval2;
1139
1140 if (requeue_pi) {
1141 /*
1142 * requeue_pi requires a pi_state, try to allocate it now
1143 * without any locks in case it fails.
1144 */
1145 if (refill_pi_state_cache())
1146 return -ENOMEM;
1147 /*
1148 * requeue_pi must wake as many tasks as it can, up to nr_wake
1149 * + nr_requeue, since it acquires the rt_mutex prior to
1150 * returning to userspace, so as to not leave the rt_mutex with
1151 * waiters and no owner. However, second and third wake-ups
1152 * cannot be predicted as they involve race conditions with the
1153 * first wake and a fault while looking up the pi_state. Both
1154 * pthread_cond_signal() and pthread_cond_broadcast() should
1155 * use nr_wake=1.
1156 */
1157 if (nr_wake != 1)
1158 return -EINVAL;
1159 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001160
Darren Hart42d35d42008-12-29 15:49:53 -08001161retry:
Darren Hart52400ba2009-04-03 13:40:49 -07001162 if (pi_state != NULL) {
1163 /*
1164 * We will have to lookup the pi_state again, so free this one
1165 * to keep the accounting correct.
1166 */
1167 free_pi_state(pi_state);
1168 pi_state = NULL;
1169 }
1170
Thomas Gleixner64d13042009-05-18 21:20:10 +02001171 ret = get_futex_key(uaddr1, fshared, &key1, VERIFY_READ);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001172 if (unlikely(ret != 0))
1173 goto out;
Thomas Gleixner521c1802009-05-20 09:02:28 +02001174 ret = get_futex_key(uaddr2, fshared, &key2,
1175 requeue_pi ? VERIFY_WRITE : VERIFY_READ);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001176 if (unlikely(ret != 0))
Darren Hart42d35d42008-12-29 15:49:53 -08001177 goto out_put_key1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001178
Ingo Molnare2970f22006-06-27 02:54:47 -07001179 hb1 = hash_futex(&key1);
1180 hb2 = hash_futex(&key2);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001181
Darren Harte4dc5b72009-03-12 00:56:13 -07001182retry_private:
Ingo Molnar8b8f3192006-07-03 00:25:05 -07001183 double_lock_hb(hb1, hb2);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001184
Ingo Molnare2970f22006-06-27 02:54:47 -07001185 if (likely(cmpval != NULL)) {
1186 u32 curval;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001187
Ingo Molnare2970f22006-06-27 02:54:47 -07001188 ret = get_futex_value_locked(&curval, uaddr1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001189
1190 if (unlikely(ret)) {
Darren Hart5eb3dc62009-03-12 00:55:52 -07001191 double_unlock_hb(hb1, hb2);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001192
Darren Harte4dc5b72009-03-12 00:56:13 -07001193 ret = get_user(curval, uaddr1);
1194 if (ret)
1195 goto out_put_keys;
1196
1197 if (!fshared)
1198 goto retry_private;
1199
Darren Hartde87fcc2009-03-12 00:55:46 -07001200 put_futex_key(fshared, &key2);
1201 put_futex_key(fshared, &key1);
Darren Harte4dc5b72009-03-12 00:56:13 -07001202 goto retry;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001203 }
Ingo Molnare2970f22006-06-27 02:54:47 -07001204 if (curval != *cmpval) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001205 ret = -EAGAIN;
1206 goto out_unlock;
1207 }
1208 }
1209
Darren Hart52400ba2009-04-03 13:40:49 -07001210 if (requeue_pi && (task_count - nr_wake < nr_requeue)) {
Darren Hartbab5bc92009-04-07 23:23:50 -07001211 /*
1212 * Attempt to acquire uaddr2 and wake the top waiter. If we
1213 * intend to requeue waiters, force setting the FUTEX_WAITERS
1214 * bit. We force this here where we are able to easily handle
1215 * faults rather in the requeue loop below.
1216 */
Darren Hart52400ba2009-04-03 13:40:49 -07001217 ret = futex_proxy_trylock_atomic(uaddr2, hb1, hb2, &key1,
Darren Hartbab5bc92009-04-07 23:23:50 -07001218 &key2, &pi_state, nr_requeue);
Darren Hart52400ba2009-04-03 13:40:49 -07001219
1220 /*
1221 * At this point the top_waiter has either taken uaddr2 or is
1222 * waiting on it. If the former, then the pi_state will not
1223 * exist yet, look it up one more time to ensure we have a
1224 * reference to it.
1225 */
1226 if (ret == 1) {
1227 WARN_ON(pi_state);
1228 task_count++;
1229 ret = get_futex_value_locked(&curval2, uaddr2);
1230 if (!ret)
1231 ret = lookup_pi_state(curval2, hb2, &key2,
1232 &pi_state);
1233 }
1234
1235 switch (ret) {
1236 case 0:
1237 break;
1238 case -EFAULT:
1239 double_unlock_hb(hb1, hb2);
1240 put_futex_key(fshared, &key2);
1241 put_futex_key(fshared, &key1);
Thomas Gleixnerd0725992009-06-11 23:15:43 +02001242 ret = fault_in_user_writeable(uaddr2);
Darren Hart52400ba2009-04-03 13:40:49 -07001243 if (!ret)
1244 goto retry;
1245 goto out;
1246 case -EAGAIN:
1247 /* The owner was exiting, try again. */
1248 double_unlock_hb(hb1, hb2);
1249 put_futex_key(fshared, &key2);
1250 put_futex_key(fshared, &key1);
1251 cond_resched();
1252 goto retry;
1253 default:
1254 goto out_unlock;
1255 }
1256 }
1257
Ingo Molnare2970f22006-06-27 02:54:47 -07001258 head1 = &hb1->chain;
Pierre Peifferec92d082007-05-09 02:35:00 -07001259 plist_for_each_entry_safe(this, next, head1, list) {
Darren Hart52400ba2009-04-03 13:40:49 -07001260 if (task_count - nr_wake >= nr_requeue)
1261 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001262
Darren Hart52400ba2009-04-03 13:40:49 -07001263 if (!match_futex(&this->key, &key1))
1264 continue;
1265
Darren Hart392741e2009-08-07 15:20:48 -07001266 /*
1267 * FUTEX_WAIT_REQEUE_PI and FUTEX_CMP_REQUEUE_PI should always
1268 * be paired with each other and no other futex ops.
1269 */
1270 if ((requeue_pi && !this->rt_waiter) ||
1271 (!requeue_pi && this->rt_waiter)) {
1272 ret = -EINVAL;
1273 break;
1274 }
Darren Hart52400ba2009-04-03 13:40:49 -07001275
1276 /*
1277 * Wake nr_wake waiters. For requeue_pi, if we acquired the
1278 * lock, we already woke the top_waiter. If not, it will be
1279 * woken by futex_unlock_pi().
1280 */
1281 if (++task_count <= nr_wake && !requeue_pi) {
1282 wake_futex(this);
1283 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001284 }
Darren Hart52400ba2009-04-03 13:40:49 -07001285
Darren Hart84bc4af2009-08-13 17:36:53 -07001286 /* Ensure we requeue to the expected futex for requeue_pi. */
1287 if (requeue_pi && !match_futex(this->requeue_pi_key, &key2)) {
1288 ret = -EINVAL;
1289 break;
1290 }
1291
Darren Hart52400ba2009-04-03 13:40:49 -07001292 /*
1293 * Requeue nr_requeue waiters and possibly one more in the case
1294 * of requeue_pi if we couldn't acquire the lock atomically.
1295 */
1296 if (requeue_pi) {
1297 /* Prepare the waiter to take the rt_mutex. */
1298 atomic_inc(&pi_state->refcount);
1299 this->pi_state = pi_state;
1300 ret = rt_mutex_start_proxy_lock(&pi_state->pi_mutex,
1301 this->rt_waiter,
1302 this->task, 1);
1303 if (ret == 1) {
1304 /* We got the lock. */
Darren Hartbeda2c72009-08-09 15:34:39 -07001305 requeue_pi_wake_futex(this, &key2, hb2);
Darren Hart52400ba2009-04-03 13:40:49 -07001306 continue;
1307 } else if (ret) {
1308 /* -EDEADLK */
1309 this->pi_state = NULL;
1310 free_pi_state(pi_state);
1311 goto out_unlock;
1312 }
1313 }
1314 requeue_futex(this, hb1, hb2, &key2);
1315 drop_count++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001316 }
1317
1318out_unlock:
Darren Hart5eb3dc62009-03-12 00:55:52 -07001319 double_unlock_hb(hb1, hb2);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001320
Darren Hartcd84a422009-04-02 14:19:38 -07001321 /*
1322 * drop_futex_key_refs() must be called outside the spinlocks. During
1323 * the requeue we moved futex_q's from the hash bucket at key1 to the
1324 * one at key2 and updated their key pointer. We no longer need to
1325 * hold the references to key1.
1326 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001327 while (--drop_count >= 0)
Rusty Russell9adef582007-05-08 00:26:42 -07001328 drop_futex_key_refs(&key1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001329
Darren Hart42d35d42008-12-29 15:49:53 -08001330out_put_keys:
Peter Zijlstra38d47c12008-09-26 19:32:20 +02001331 put_futex_key(fshared, &key2);
Darren Hart42d35d42008-12-29 15:49:53 -08001332out_put_key1:
Peter Zijlstra38d47c12008-09-26 19:32:20 +02001333 put_futex_key(fshared, &key1);
Darren Hart42d35d42008-12-29 15:49:53 -08001334out:
Darren Hart52400ba2009-04-03 13:40:49 -07001335 if (pi_state != NULL)
1336 free_pi_state(pi_state);
1337 return ret ? ret : task_count;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001338}
1339
1340/* The key must be already stored in q->key. */
Eric Sesterhenn82af7ac2008-01-25 10:40:46 +01001341static inline struct futex_hash_bucket *queue_lock(struct futex_q *q)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001342{
Ingo Molnare2970f22006-06-27 02:54:47 -07001343 struct futex_hash_bucket *hb;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001344
Rusty Russell9adef582007-05-08 00:26:42 -07001345 get_futex_key_refs(&q->key);
Ingo Molnare2970f22006-06-27 02:54:47 -07001346 hb = hash_futex(&q->key);
1347 q->lock_ptr = &hb->lock;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001348
Ingo Molnare2970f22006-06-27 02:54:47 -07001349 spin_lock(&hb->lock);
1350 return hb;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001351}
1352
Eric Sesterhenn82af7ac2008-01-25 10:40:46 +01001353static inline void queue_me(struct futex_q *q, struct futex_hash_bucket *hb)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001354{
Pierre Peifferec92d082007-05-09 02:35:00 -07001355 int prio;
1356
1357 /*
1358 * The priority used to register this element is
1359 * - either the real thread-priority for the real-time threads
1360 * (i.e. threads with a priority lower than MAX_RT_PRIO)
1361 * - or MAX_RT_PRIO for non-RT threads.
1362 * Thus, all RT-threads are woken first in priority order, and
1363 * the others are woken last, in FIFO order.
1364 */
1365 prio = min(current->normal_prio, MAX_RT_PRIO);
1366
1367 plist_node_init(&q->list, prio);
1368#ifdef CONFIG_DEBUG_PI_LIST
1369 q->list.plist.lock = &hb->lock;
1370#endif
1371 plist_add(&q->list, &hb->chain);
Ingo Molnarc87e2832006-06-27 02:54:58 -07001372 q->task = current;
Ingo Molnare2970f22006-06-27 02:54:47 -07001373 spin_unlock(&hb->lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001374}
1375
1376static inline void
Ingo Molnare2970f22006-06-27 02:54:47 -07001377queue_unlock(struct futex_q *q, struct futex_hash_bucket *hb)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001378{
Ingo Molnare2970f22006-06-27 02:54:47 -07001379 spin_unlock(&hb->lock);
Rusty Russell9adef582007-05-08 00:26:42 -07001380 drop_futex_key_refs(&q->key);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001381}
1382
1383/*
1384 * queue_me and unqueue_me must be called as a pair, each
1385 * exactly once. They are called with the hashed spinlock held.
1386 */
1387
Linus Torvalds1da177e2005-04-16 15:20:36 -07001388/* Return 1 if we were still queued (ie. 0 means we were woken) */
1389static int unqueue_me(struct futex_q *q)
1390{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001391 spinlock_t *lock_ptr;
Ingo Molnare2970f22006-06-27 02:54:47 -07001392 int ret = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001393
1394 /* In the common case we don't take the spinlock, which is nice. */
Darren Hart42d35d42008-12-29 15:49:53 -08001395retry:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001396 lock_ptr = q->lock_ptr;
Christian Borntraegere91467e2006-08-05 12:13:52 -07001397 barrier();
Stephen Hemmingerc80544d2007-10-18 03:07:05 -07001398 if (lock_ptr != NULL) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001399 spin_lock(lock_ptr);
1400 /*
1401 * q->lock_ptr can change between reading it and
1402 * spin_lock(), causing us to take the wrong lock. This
1403 * corrects the race condition.
1404 *
1405 * Reasoning goes like this: if we have the wrong lock,
1406 * q->lock_ptr must have changed (maybe several times)
1407 * between reading it and the spin_lock(). It can
1408 * change again after the spin_lock() but only if it was
1409 * already changed before the spin_lock(). It cannot,
1410 * however, change back to the original value. Therefore
1411 * we can detect whether we acquired the correct lock.
1412 */
1413 if (unlikely(lock_ptr != q->lock_ptr)) {
1414 spin_unlock(lock_ptr);
1415 goto retry;
1416 }
Pierre Peifferec92d082007-05-09 02:35:00 -07001417 WARN_ON(plist_node_empty(&q->list));
1418 plist_del(&q->list, &q->list.plist);
Ingo Molnarc87e2832006-06-27 02:54:58 -07001419
1420 BUG_ON(q->pi_state);
1421
Linus Torvalds1da177e2005-04-16 15:20:36 -07001422 spin_unlock(lock_ptr);
1423 ret = 1;
1424 }
1425
Rusty Russell9adef582007-05-08 00:26:42 -07001426 drop_futex_key_refs(&q->key);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001427 return ret;
1428}
1429
Ingo Molnarc87e2832006-06-27 02:54:58 -07001430/*
1431 * PI futexes can not be requeued and must remove themself from the
Pierre Peifferd0aa7a72007-05-09 02:35:02 -07001432 * hash bucket. The hash bucket lock (i.e. lock_ptr) is held on entry
1433 * and dropped here.
Ingo Molnarc87e2832006-06-27 02:54:58 -07001434 */
Pierre Peifferd0aa7a72007-05-09 02:35:02 -07001435static void unqueue_me_pi(struct futex_q *q)
Ingo Molnarc87e2832006-06-27 02:54:58 -07001436{
Pierre Peifferec92d082007-05-09 02:35:00 -07001437 WARN_ON(plist_node_empty(&q->list));
1438 plist_del(&q->list, &q->list.plist);
Ingo Molnarc87e2832006-06-27 02:54:58 -07001439
1440 BUG_ON(!q->pi_state);
1441 free_pi_state(q->pi_state);
1442 q->pi_state = NULL;
1443
Pierre Peifferd0aa7a72007-05-09 02:35:02 -07001444 spin_unlock(q->lock_ptr);
Ingo Molnarc87e2832006-06-27 02:54:58 -07001445
Rusty Russell9adef582007-05-08 00:26:42 -07001446 drop_futex_key_refs(&q->key);
Ingo Molnarc87e2832006-06-27 02:54:58 -07001447}
1448
Pierre Peifferd0aa7a72007-05-09 02:35:02 -07001449/*
Thomas Gleixnercdf71a12008-01-08 19:47:38 +01001450 * Fixup the pi_state owner with the new owner.
Pierre Peifferd0aa7a72007-05-09 02:35:02 -07001451 *
Alexey Kuznetsov778e9a92007-06-08 13:47:00 -07001452 * Must be called with hash bucket lock held and mm->sem held for non
1453 * private futexes.
Pierre Peifferd0aa7a72007-05-09 02:35:02 -07001454 */
Alexey Kuznetsov778e9a92007-06-08 13:47:00 -07001455static int fixup_pi_state_owner(u32 __user *uaddr, struct futex_q *q,
Peter Zijlstrac2f9f202008-09-26 19:32:23 +02001456 struct task_struct *newowner, int fshared)
Pierre Peifferd0aa7a72007-05-09 02:35:02 -07001457{
Thomas Gleixnercdf71a12008-01-08 19:47:38 +01001458 u32 newtid = task_pid_vnr(newowner) | FUTEX_WAITERS;
Pierre Peifferd0aa7a72007-05-09 02:35:02 -07001459 struct futex_pi_state *pi_state = q->pi_state;
Thomas Gleixner1b7558e2008-06-23 11:21:58 +02001460 struct task_struct *oldowner = pi_state->owner;
Pierre Peifferd0aa7a72007-05-09 02:35:02 -07001461 u32 uval, curval, newval;
Darren Harte4dc5b72009-03-12 00:56:13 -07001462 int ret;
Pierre Peifferd0aa7a72007-05-09 02:35:02 -07001463
1464 /* Owner died? */
Thomas Gleixner1b7558e2008-06-23 11:21:58 +02001465 if (!pi_state->owner)
1466 newtid |= FUTEX_OWNER_DIED;
1467
1468 /*
1469 * We are here either because we stole the rtmutex from the
1470 * pending owner or we are the pending owner which failed to
1471 * get the rtmutex. We have to replace the pending owner TID
1472 * in the user space variable. This must be atomic as we have
1473 * to preserve the owner died bit here.
1474 *
Darren Hartb2d09942009-03-12 00:55:37 -07001475 * Note: We write the user space value _before_ changing the pi_state
1476 * because we can fault here. Imagine swapped out pages or a fork
1477 * that marked all the anonymous memory readonly for cow.
Thomas Gleixner1b7558e2008-06-23 11:21:58 +02001478 *
1479 * Modifying pi_state _before_ the user space value would
1480 * leave the pi_state in an inconsistent state when we fault
1481 * here, because we need to drop the hash bucket lock to
1482 * handle the fault. This might be observed in the PID check
1483 * in lookup_pi_state.
1484 */
1485retry:
1486 if (get_futex_value_locked(&uval, uaddr))
1487 goto handle_fault;
1488
1489 while (1) {
1490 newval = (uval & FUTEX_OWNER_DIED) | newtid;
1491
1492 curval = cmpxchg_futex_value_locked(uaddr, uval, newval);
1493
1494 if (curval == -EFAULT)
1495 goto handle_fault;
1496 if (curval == uval)
1497 break;
1498 uval = curval;
1499 }
1500
1501 /*
1502 * We fixed up user space. Now we need to fix the pi_state
1503 * itself.
1504 */
Pierre Peifferd0aa7a72007-05-09 02:35:02 -07001505 if (pi_state->owner != NULL) {
1506 spin_lock_irq(&pi_state->owner->pi_lock);
1507 WARN_ON(list_empty(&pi_state->list));
1508 list_del_init(&pi_state->list);
1509 spin_unlock_irq(&pi_state->owner->pi_lock);
Thomas Gleixner1b7558e2008-06-23 11:21:58 +02001510 }
Pierre Peifferd0aa7a72007-05-09 02:35:02 -07001511
Thomas Gleixnercdf71a12008-01-08 19:47:38 +01001512 pi_state->owner = newowner;
Pierre Peifferd0aa7a72007-05-09 02:35:02 -07001513
Thomas Gleixnercdf71a12008-01-08 19:47:38 +01001514 spin_lock_irq(&newowner->pi_lock);
Pierre Peifferd0aa7a72007-05-09 02:35:02 -07001515 WARN_ON(!list_empty(&pi_state->list));
Thomas Gleixnercdf71a12008-01-08 19:47:38 +01001516 list_add(&pi_state->list, &newowner->pi_state_list);
1517 spin_unlock_irq(&newowner->pi_lock);
Thomas Gleixner1b7558e2008-06-23 11:21:58 +02001518 return 0;
Pierre Peifferd0aa7a72007-05-09 02:35:02 -07001519
Pierre Peifferd0aa7a72007-05-09 02:35:02 -07001520 /*
Thomas Gleixner1b7558e2008-06-23 11:21:58 +02001521 * To handle the page fault we need to drop the hash bucket
1522 * lock here. That gives the other task (either the pending
1523 * owner itself or the task which stole the rtmutex) the
1524 * chance to try the fixup of the pi_state. So once we are
1525 * back from handling the fault we need to check the pi_state
1526 * after reacquiring the hash bucket lock and before trying to
1527 * do another fixup. When the fixup has been done already we
1528 * simply return.
Pierre Peifferd0aa7a72007-05-09 02:35:02 -07001529 */
Thomas Gleixner1b7558e2008-06-23 11:21:58 +02001530handle_fault:
1531 spin_unlock(q->lock_ptr);
Alexey Kuznetsov778e9a92007-06-08 13:47:00 -07001532
Thomas Gleixnerd0725992009-06-11 23:15:43 +02001533 ret = fault_in_user_writeable(uaddr);
Alexey Kuznetsov778e9a92007-06-08 13:47:00 -07001534
Thomas Gleixner1b7558e2008-06-23 11:21:58 +02001535 spin_lock(q->lock_ptr);
Alexey Kuznetsov778e9a92007-06-08 13:47:00 -07001536
Thomas Gleixner1b7558e2008-06-23 11:21:58 +02001537 /*
1538 * Check if someone else fixed it for us:
1539 */
1540 if (pi_state->owner != oldowner)
1541 return 0;
1542
1543 if (ret)
1544 return ret;
1545
1546 goto retry;
Pierre Peifferd0aa7a72007-05-09 02:35:02 -07001547}
1548
Eric Dumazet34f01cc2007-05-09 02:35:04 -07001549/*
1550 * In case we must use restart_block to restart a futex_wait,
Steven Rostedtce6bd422007-12-05 15:46:09 +01001551 * we encode in the 'flags' shared capability
Eric Dumazet34f01cc2007-05-09 02:35:04 -07001552 */
Thomas Gleixner1acdac12008-11-20 10:02:53 -08001553#define FLAGS_SHARED 0x01
1554#define FLAGS_CLOCKRT 0x02
Darren Harta72188d2009-04-03 13:40:22 -07001555#define FLAGS_HAS_TIMEOUT 0x04
Eric Dumazet34f01cc2007-05-09 02:35:04 -07001556
Nick Piggin72c1bbf2007-05-08 00:26:43 -07001557static long futex_wait_restart(struct restart_block *restart);
Thomas Gleixner36cf3b52007-07-15 23:41:20 -07001558
Darren Hartca5f9522009-04-03 13:39:33 -07001559/**
Darren Hartdd973992009-04-03 13:40:02 -07001560 * fixup_owner() - Post lock pi_state and corner case management
1561 * @uaddr: user address of the futex
1562 * @fshared: whether the futex is shared (1) or not (0)
1563 * @q: futex_q (contains pi_state and access to the rt_mutex)
1564 * @locked: if the attempt to take the rt_mutex succeeded (1) or not (0)
1565 *
1566 * After attempting to lock an rt_mutex, this function is called to cleanup
1567 * the pi_state owner as well as handle race conditions that may allow us to
1568 * acquire the lock. Must be called with the hb lock held.
1569 *
1570 * Returns:
1571 * 1 - success, lock taken
1572 * 0 - success, lock not taken
1573 * <0 - on error (-EFAULT)
1574 */
1575static int fixup_owner(u32 __user *uaddr, int fshared, struct futex_q *q,
1576 int locked)
1577{
1578 struct task_struct *owner;
1579 int ret = 0;
1580
1581 if (locked) {
1582 /*
1583 * Got the lock. We might not be the anticipated owner if we
1584 * did a lock-steal - fix up the PI-state in that case:
1585 */
1586 if (q->pi_state->owner != current)
1587 ret = fixup_pi_state_owner(uaddr, q, current, fshared);
1588 goto out;
1589 }
1590
1591 /*
1592 * Catch the rare case, where the lock was released when we were on the
1593 * way back before we locked the hash bucket.
1594 */
1595 if (q->pi_state->owner == current) {
1596 /*
1597 * Try to get the rt_mutex now. This might fail as some other
1598 * task acquired the rt_mutex after we removed ourself from the
1599 * rt_mutex waiters list.
1600 */
1601 if (rt_mutex_trylock(&q->pi_state->pi_mutex)) {
1602 locked = 1;
1603 goto out;
1604 }
1605
1606 /*
1607 * pi_state is incorrect, some other task did a lock steal and
1608 * we returned due to timeout or signal without taking the
1609 * rt_mutex. Too late. We can access the rt_mutex_owner without
1610 * locking, as the other task is now blocked on the hash bucket
1611 * lock. Fix the state up.
1612 */
1613 owner = rt_mutex_owner(&q->pi_state->pi_mutex);
1614 ret = fixup_pi_state_owner(uaddr, q, owner, fshared);
1615 goto out;
1616 }
1617
1618 /*
1619 * Paranoia check. If we did not take the lock, then we should not be
1620 * the owner, nor the pending owner, of the rt_mutex.
1621 */
1622 if (rt_mutex_owner(&q->pi_state->pi_mutex) == current)
1623 printk(KERN_ERR "fixup_owner: ret = %d pi-mutex: %p "
1624 "pi-state %p\n", ret,
1625 q->pi_state->pi_mutex.owner,
1626 q->pi_state->owner);
1627
1628out:
1629 return ret ? ret : locked;
1630}
1631
1632/**
Darren Hartca5f9522009-04-03 13:39:33 -07001633 * futex_wait_queue_me() - queue_me() and wait for wakeup, timeout, or signal
1634 * @hb: the futex hash bucket, must be locked by the caller
1635 * @q: the futex_q to queue up on
1636 * @timeout: the prepared hrtimer_sleeper, or null for no timeout
Darren Hartca5f9522009-04-03 13:39:33 -07001637 */
1638static void futex_wait_queue_me(struct futex_hash_bucket *hb, struct futex_q *q,
Thomas Gleixnerf1a11e02009-05-05 19:21:40 +02001639 struct hrtimer_sleeper *timeout)
Darren Hartca5f9522009-04-03 13:39:33 -07001640{
1641 queue_me(q, hb);
1642
1643 /*
1644 * There might have been scheduling since the queue_me(), as we
1645 * cannot hold a spinlock across the get_user() in case it
1646 * faults, and we cannot just set TASK_INTERRUPTIBLE state when
Thomas Gleixnerf1a11e02009-05-05 19:21:40 +02001647 * queueing ourselves into the futex hash. This code thus has to
Darren Hartca5f9522009-04-03 13:39:33 -07001648 * rely on the futex_wake() code removing us from hash when it
1649 * wakes us up.
1650 */
Thomas Gleixnerf1a11e02009-05-05 19:21:40 +02001651 set_current_state(TASK_INTERRUPTIBLE);
Darren Hartca5f9522009-04-03 13:39:33 -07001652
1653 /* Arm the timer */
1654 if (timeout) {
1655 hrtimer_start_expires(&timeout->timer, HRTIMER_MODE_ABS);
1656 if (!hrtimer_active(&timeout->timer))
1657 timeout->task = NULL;
1658 }
1659
1660 /*
1661 * !plist_node_empty() is safe here without any lock.
1662 * q.lock_ptr != 0 is not safe, because of ordering against wakeup.
1663 */
1664 if (likely(!plist_node_empty(&q->list))) {
1665 /*
1666 * If the timer has already expired, current will already be
1667 * flagged for rescheduling. Only call schedule if there
1668 * is no timeout, or if it has yet to expire.
1669 */
1670 if (!timeout || timeout->task)
1671 schedule();
1672 }
1673 __set_current_state(TASK_RUNNING);
1674}
1675
Darren Hartf8010732009-04-03 13:40:40 -07001676/**
1677 * futex_wait_setup() - Prepare to wait on a futex
1678 * @uaddr: the futex userspace address
1679 * @val: the expected value
1680 * @fshared: whether the futex is shared (1) or not (0)
1681 * @q: the associated futex_q
1682 * @hb: storage for hash_bucket pointer to be returned to caller
1683 *
1684 * Setup the futex_q and locate the hash_bucket. Get the futex value and
1685 * compare it with the expected value. Handle atomic faults internally.
1686 * Return with the hb lock held and a q.key reference on success, and unlocked
1687 * with no q.key reference on failure.
1688 *
1689 * Returns:
1690 * 0 - uaddr contains val and hb has been locked
1691 * <1 - -EFAULT or -EWOULDBLOCK (uaddr does not contain val) and hb is unlcoked
1692 */
1693static int futex_wait_setup(u32 __user *uaddr, u32 val, int fshared,
1694 struct futex_q *q, struct futex_hash_bucket **hb)
1695{
1696 u32 uval;
1697 int ret;
1698
1699 /*
1700 * Access the page AFTER the hash-bucket is locked.
1701 * Order is important:
1702 *
1703 * Userspace waiter: val = var; if (cond(val)) futex_wait(&var, val);
1704 * Userspace waker: if (cond(var)) { var = new; futex_wake(&var); }
1705 *
1706 * The basic logical guarantee of a futex is that it blocks ONLY
1707 * if cond(var) is known to be true at the time of blocking, for
1708 * any cond. If we queued after testing *uaddr, that would open
1709 * a race condition where we could block indefinitely with
1710 * cond(var) false, which would violate the guarantee.
1711 *
1712 * A consequence is that futex_wait() can return zero and absorb
1713 * a wakeup when *uaddr != val on entry to the syscall. This is
1714 * rare, but normal.
1715 */
1716retry:
1717 q->key = FUTEX_KEY_INIT;
Thomas Gleixner521c1802009-05-20 09:02:28 +02001718 ret = get_futex_key(uaddr, fshared, &q->key, VERIFY_READ);
Darren Hartf8010732009-04-03 13:40:40 -07001719 if (unlikely(ret != 0))
Darren Harta5a2a0c2009-04-10 09:50:05 -07001720 return ret;
Darren Hartf8010732009-04-03 13:40:40 -07001721
1722retry_private:
1723 *hb = queue_lock(q);
1724
1725 ret = get_futex_value_locked(&uval, uaddr);
1726
1727 if (ret) {
1728 queue_unlock(q, *hb);
1729
1730 ret = get_user(uval, uaddr);
1731 if (ret)
1732 goto out;
1733
1734 if (!fshared)
1735 goto retry_private;
1736
1737 put_futex_key(fshared, &q->key);
1738 goto retry;
1739 }
1740
1741 if (uval != val) {
1742 queue_unlock(q, *hb);
1743 ret = -EWOULDBLOCK;
1744 }
1745
1746out:
1747 if (ret)
1748 put_futex_key(fshared, &q->key);
1749 return ret;
1750}
1751
Peter Zijlstrac2f9f202008-09-26 19:32:23 +02001752static int futex_wait(u32 __user *uaddr, int fshared,
Thomas Gleixner1acdac12008-11-20 10:02:53 -08001753 u32 val, ktime_t *abs_time, u32 bitset, int clockrt)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001754{
Darren Hartca5f9522009-04-03 13:39:33 -07001755 struct hrtimer_sleeper timeout, *to = NULL;
Peter Zijlstra2fff78c2009-02-11 18:10:10 +01001756 struct restart_block *restart;
Ingo Molnare2970f22006-06-27 02:54:47 -07001757 struct futex_hash_bucket *hb;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001758 struct futex_q q;
Ingo Molnare2970f22006-06-27 02:54:47 -07001759 int ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001760
Thomas Gleixnercd689982008-02-01 17:45:14 +01001761 if (!bitset)
1762 return -EINVAL;
1763
Ingo Molnarc87e2832006-06-27 02:54:58 -07001764 q.pi_state = NULL;
Thomas Gleixnercd689982008-02-01 17:45:14 +01001765 q.bitset = bitset;
Darren Hart52400ba2009-04-03 13:40:49 -07001766 q.rt_waiter = NULL;
Darren Hart84bc4af2009-08-13 17:36:53 -07001767 q.requeue_pi_key = NULL;
Darren Hartca5f9522009-04-03 13:39:33 -07001768
1769 if (abs_time) {
1770 to = &timeout;
1771
1772 hrtimer_init_on_stack(&to->timer, clockrt ? CLOCK_REALTIME :
1773 CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
1774 hrtimer_init_sleeper(to, current);
1775 hrtimer_set_expires_range_ns(&to->timer, *abs_time,
1776 current->timer_slack_ns);
1777 }
1778
Darren Hartf8010732009-04-03 13:40:40 -07001779 /* Prepare to wait on uaddr. */
1780 ret = futex_wait_setup(uaddr, val, fshared, &q, &hb);
1781 if (ret)
Darren Hart42d35d42008-12-29 15:49:53 -08001782 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001783
Darren Hartca5f9522009-04-03 13:39:33 -07001784 /* queue_me and wait for wakeup, timeout, or a signal. */
Thomas Gleixnerf1a11e02009-05-05 19:21:40 +02001785 futex_wait_queue_me(hb, &q, to);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001786
1787 /* If we were woken (and unqueued), we succeeded, whatever. */
Peter Zijlstra2fff78c2009-02-11 18:10:10 +01001788 ret = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001789 if (!unqueue_me(&q))
Peter Zijlstra2fff78c2009-02-11 18:10:10 +01001790 goto out_put_key;
1791 ret = -ETIMEDOUT;
Darren Hartca5f9522009-04-03 13:39:33 -07001792 if (to && !to->task)
Peter Zijlstra2fff78c2009-02-11 18:10:10 +01001793 goto out_put_key;
Nick Piggin72c1bbf2007-05-08 00:26:43 -07001794
Ingo Molnare2970f22006-06-27 02:54:47 -07001795 /*
1796 * We expect signal_pending(current), but another thread may
1797 * have handled it for us already.
1798 */
Peter Zijlstra2fff78c2009-02-11 18:10:10 +01001799 ret = -ERESTARTSYS;
Pierre Peifferc19384b2007-05-09 02:35:02 -07001800 if (!abs_time)
Peter Zijlstra2fff78c2009-02-11 18:10:10 +01001801 goto out_put_key;
Steven Rostedtce6bd422007-12-05 15:46:09 +01001802
Peter Zijlstra2fff78c2009-02-11 18:10:10 +01001803 restart = &current_thread_info()->restart_block;
1804 restart->fn = futex_wait_restart;
1805 restart->futex.uaddr = (u32 *)uaddr;
1806 restart->futex.val = val;
1807 restart->futex.time = abs_time->tv64;
1808 restart->futex.bitset = bitset;
Darren Harta72188d2009-04-03 13:40:22 -07001809 restart->futex.flags = FLAGS_HAS_TIMEOUT;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001810
Peter Zijlstra2fff78c2009-02-11 18:10:10 +01001811 if (fshared)
1812 restart->futex.flags |= FLAGS_SHARED;
1813 if (clockrt)
1814 restart->futex.flags |= FLAGS_CLOCKRT;
1815
1816 ret = -ERESTART_RESTARTBLOCK;
1817
1818out_put_key:
Peter Zijlstra38d47c12008-09-26 19:32:20 +02001819 put_futex_key(fshared, &q.key);
Darren Hart42d35d42008-12-29 15:49:53 -08001820out:
Darren Hartca5f9522009-04-03 13:39:33 -07001821 if (to) {
1822 hrtimer_cancel(&to->timer);
1823 destroy_hrtimer_on_stack(&to->timer);
1824 }
Ingo Molnarc87e2832006-06-27 02:54:58 -07001825 return ret;
1826}
1827
Nick Piggin72c1bbf2007-05-08 00:26:43 -07001828
1829static long futex_wait_restart(struct restart_block *restart)
1830{
Steven Rostedtce6bd422007-12-05 15:46:09 +01001831 u32 __user *uaddr = (u32 __user *)restart->futex.uaddr;
Peter Zijlstrac2f9f202008-09-26 19:32:23 +02001832 int fshared = 0;
Darren Harta72188d2009-04-03 13:40:22 -07001833 ktime_t t, *tp = NULL;
Nick Piggin72c1bbf2007-05-08 00:26:43 -07001834
Darren Harta72188d2009-04-03 13:40:22 -07001835 if (restart->futex.flags & FLAGS_HAS_TIMEOUT) {
1836 t.tv64 = restart->futex.time;
1837 tp = &t;
1838 }
Nick Piggin72c1bbf2007-05-08 00:26:43 -07001839 restart->fn = do_no_restart_syscall;
Steven Rostedtce6bd422007-12-05 15:46:09 +01001840 if (restart->futex.flags & FLAGS_SHARED)
Peter Zijlstrac2f9f202008-09-26 19:32:23 +02001841 fshared = 1;
Darren Harta72188d2009-04-03 13:40:22 -07001842 return (long)futex_wait(uaddr, fshared, restart->futex.val, tp,
Thomas Gleixner1acdac12008-11-20 10:02:53 -08001843 restart->futex.bitset,
1844 restart->futex.flags & FLAGS_CLOCKRT);
Nick Piggin72c1bbf2007-05-08 00:26:43 -07001845}
1846
1847
Ingo Molnarc87e2832006-06-27 02:54:58 -07001848/*
1849 * Userspace tried a 0 -> TID atomic transition of the futex value
1850 * and failed. The kernel side here does the whole locking operation:
1851 * if there are waiters then it will block, it does PI, etc. (Due to
1852 * races the kernel might see a 0 value of the futex too.)
1853 */
Peter Zijlstrac2f9f202008-09-26 19:32:23 +02001854static int futex_lock_pi(u32 __user *uaddr, int fshared,
Eric Dumazet34f01cc2007-05-09 02:35:04 -07001855 int detect, ktime_t *time, int trylock)
Ingo Molnarc87e2832006-06-27 02:54:58 -07001856{
Thomas Gleixnerc5780e92006-09-08 09:47:15 -07001857 struct hrtimer_sleeper timeout, *to = NULL;
Ingo Molnarc87e2832006-06-27 02:54:58 -07001858 struct futex_hash_bucket *hb;
Ingo Molnarc87e2832006-06-27 02:54:58 -07001859 struct futex_q q;
Darren Hartdd973992009-04-03 13:40:02 -07001860 int res, ret;
Ingo Molnarc87e2832006-06-27 02:54:58 -07001861
1862 if (refill_pi_state_cache())
1863 return -ENOMEM;
1864
Pierre Peifferc19384b2007-05-09 02:35:02 -07001865 if (time) {
Thomas Gleixnerc5780e92006-09-08 09:47:15 -07001866 to = &timeout;
Thomas Gleixner237fc6e2008-04-30 00:55:04 -07001867 hrtimer_init_on_stack(&to->timer, CLOCK_REALTIME,
1868 HRTIMER_MODE_ABS);
Thomas Gleixnerc5780e92006-09-08 09:47:15 -07001869 hrtimer_init_sleeper(to, current);
Arjan van de Vencc584b22008-09-01 15:02:30 -07001870 hrtimer_set_expires(&to->timer, *time);
Thomas Gleixnerc5780e92006-09-08 09:47:15 -07001871 }
1872
Ingo Molnarc87e2832006-06-27 02:54:58 -07001873 q.pi_state = NULL;
Darren Hart52400ba2009-04-03 13:40:49 -07001874 q.rt_waiter = NULL;
Darren Hart84bc4af2009-08-13 17:36:53 -07001875 q.requeue_pi_key = NULL;
Darren Hart42d35d42008-12-29 15:49:53 -08001876retry:
Peter Zijlstra38d47c12008-09-26 19:32:20 +02001877 q.key = FUTEX_KEY_INIT;
Thomas Gleixner64d13042009-05-18 21:20:10 +02001878 ret = get_futex_key(uaddr, fshared, &q.key, VERIFY_WRITE);
Ingo Molnarc87e2832006-06-27 02:54:58 -07001879 if (unlikely(ret != 0))
Darren Hart42d35d42008-12-29 15:49:53 -08001880 goto out;
Ingo Molnarc87e2832006-06-27 02:54:58 -07001881
Darren Harte4dc5b72009-03-12 00:56:13 -07001882retry_private:
Eric Sesterhenn82af7ac2008-01-25 10:40:46 +01001883 hb = queue_lock(&q);
Ingo Molnarc87e2832006-06-27 02:54:58 -07001884
Darren Hartbab5bc92009-04-07 23:23:50 -07001885 ret = futex_lock_pi_atomic(uaddr, hb, &q.key, &q.pi_state, current, 0);
Ingo Molnarc87e2832006-06-27 02:54:58 -07001886 if (unlikely(ret)) {
Alexey Kuznetsov778e9a92007-06-08 13:47:00 -07001887 switch (ret) {
Darren Hart1a520842009-04-03 13:39:52 -07001888 case 1:
1889 /* We got the lock. */
1890 ret = 0;
1891 goto out_unlock_put_key;
1892 case -EFAULT:
1893 goto uaddr_faulted;
Alexey Kuznetsov778e9a92007-06-08 13:47:00 -07001894 case -EAGAIN:
1895 /*
1896 * Task is exiting and we just wait for the
1897 * exit to complete.
1898 */
1899 queue_unlock(&q, hb);
Darren Hartde87fcc2009-03-12 00:55:46 -07001900 put_futex_key(fshared, &q.key);
Alexey Kuznetsov778e9a92007-06-08 13:47:00 -07001901 cond_resched();
1902 goto retry;
Alexey Kuznetsov778e9a92007-06-08 13:47:00 -07001903 default:
Darren Hart42d35d42008-12-29 15:49:53 -08001904 goto out_unlock_put_key;
Ingo Molnarc87e2832006-06-27 02:54:58 -07001905 }
Ingo Molnarc87e2832006-06-27 02:54:58 -07001906 }
1907
1908 /*
1909 * Only actually queue now that the atomic ops are done:
1910 */
Eric Sesterhenn82af7ac2008-01-25 10:40:46 +01001911 queue_me(&q, hb);
Ingo Molnarc87e2832006-06-27 02:54:58 -07001912
Ingo Molnarc87e2832006-06-27 02:54:58 -07001913 WARN_ON(!q.pi_state);
1914 /*
1915 * Block on the PI mutex:
1916 */
1917 if (!trylock)
1918 ret = rt_mutex_timed_lock(&q.pi_state->pi_mutex, to, 1);
1919 else {
1920 ret = rt_mutex_trylock(&q.pi_state->pi_mutex);
1921 /* Fixup the trylock return value: */
1922 ret = ret ? 0 : -EWOULDBLOCK;
1923 }
1924
Vernon Mauerya99e4e42006-07-01 04:35:42 -07001925 spin_lock(q.lock_ptr);
Darren Hartdd973992009-04-03 13:40:02 -07001926 /*
1927 * Fixup the pi_state owner and possibly acquire the lock if we
1928 * haven't already.
1929 */
1930 res = fixup_owner(uaddr, fshared, &q, !ret);
1931 /*
1932 * If fixup_owner() returned an error, proprogate that. If it acquired
1933 * the lock, clear our -ETIMEDOUT or -EINTR.
1934 */
1935 if (res)
1936 ret = (res < 0) ? res : 0;
Ingo Molnarc87e2832006-06-27 02:54:58 -07001937
Darren Harte8f63862009-03-12 00:56:06 -07001938 /*
Darren Hartdd973992009-04-03 13:40:02 -07001939 * If fixup_owner() faulted and was unable to handle the fault, unlock
1940 * it and return the fault to userspace.
Darren Harte8f63862009-03-12 00:56:06 -07001941 */
1942 if (ret && (rt_mutex_owner(&q.pi_state->pi_mutex) == current))
1943 rt_mutex_unlock(&q.pi_state->pi_mutex);
1944
Alexey Kuznetsov778e9a92007-06-08 13:47:00 -07001945 /* Unqueue and drop the lock */
1946 unqueue_me_pi(&q);
Ingo Molnarc87e2832006-06-27 02:54:58 -07001947
Darren Hartdd973992009-04-03 13:40:02 -07001948 goto out;
Ingo Molnarc87e2832006-06-27 02:54:58 -07001949
Darren Hart42d35d42008-12-29 15:49:53 -08001950out_unlock_put_key:
Ingo Molnarc87e2832006-06-27 02:54:58 -07001951 queue_unlock(&q, hb);
1952
Darren Hart42d35d42008-12-29 15:49:53 -08001953out_put_key:
Peter Zijlstra38d47c12008-09-26 19:32:20 +02001954 put_futex_key(fshared, &q.key);
Darren Hart42d35d42008-12-29 15:49:53 -08001955out:
Thomas Gleixner237fc6e2008-04-30 00:55:04 -07001956 if (to)
1957 destroy_hrtimer_on_stack(&to->timer);
Darren Hartdd973992009-04-03 13:40:02 -07001958 return ret != -EINTR ? ret : -ERESTARTNOINTR;
Ingo Molnarc87e2832006-06-27 02:54:58 -07001959
Darren Hart42d35d42008-12-29 15:49:53 -08001960uaddr_faulted:
Alexey Kuznetsov778e9a92007-06-08 13:47:00 -07001961 queue_unlock(&q, hb);
1962
Thomas Gleixnerd0725992009-06-11 23:15:43 +02001963 ret = fault_in_user_writeable(uaddr);
Darren Harte4dc5b72009-03-12 00:56:13 -07001964 if (ret)
1965 goto out_put_key;
Ingo Molnarc87e2832006-06-27 02:54:58 -07001966
Darren Harte4dc5b72009-03-12 00:56:13 -07001967 if (!fshared)
1968 goto retry_private;
1969
1970 put_futex_key(fshared, &q.key);
1971 goto retry;
Ingo Molnarc87e2832006-06-27 02:54:58 -07001972}
1973
1974/*
Ingo Molnarc87e2832006-06-27 02:54:58 -07001975 * Userspace attempted a TID -> 0 atomic transition, and failed.
1976 * This is the in-kernel slowpath: we look up the PI state (if any),
1977 * and do the rt-mutex unlock.
1978 */
Peter Zijlstrac2f9f202008-09-26 19:32:23 +02001979static int futex_unlock_pi(u32 __user *uaddr, int fshared)
Ingo Molnarc87e2832006-06-27 02:54:58 -07001980{
1981 struct futex_hash_bucket *hb;
1982 struct futex_q *this, *next;
1983 u32 uval;
Pierre Peifferec92d082007-05-09 02:35:00 -07001984 struct plist_head *head;
Peter Zijlstra38d47c12008-09-26 19:32:20 +02001985 union futex_key key = FUTEX_KEY_INIT;
Darren Harte4dc5b72009-03-12 00:56:13 -07001986 int ret;
Ingo Molnarc87e2832006-06-27 02:54:58 -07001987
1988retry:
1989 if (get_user(uval, uaddr))
1990 return -EFAULT;
1991 /*
1992 * We release only a lock we actually own:
1993 */
Pavel Emelyanovb4888932007-10-18 23:40:14 -07001994 if ((uval & FUTEX_TID_MASK) != task_pid_vnr(current))
Ingo Molnarc87e2832006-06-27 02:54:58 -07001995 return -EPERM;
Ingo Molnarc87e2832006-06-27 02:54:58 -07001996
Thomas Gleixner64d13042009-05-18 21:20:10 +02001997 ret = get_futex_key(uaddr, fshared, &key, VERIFY_WRITE);
Ingo Molnarc87e2832006-06-27 02:54:58 -07001998 if (unlikely(ret != 0))
1999 goto out;
2000
2001 hb = hash_futex(&key);
2002 spin_lock(&hb->lock);
2003
Ingo Molnarc87e2832006-06-27 02:54:58 -07002004 /*
2005 * To avoid races, try to do the TID -> 0 atomic transition
2006 * again. If it succeeds then we can return without waking
2007 * anyone else up:
2008 */
Thomas Gleixner36cf3b52007-07-15 23:41:20 -07002009 if (!(uval & FUTEX_OWNER_DIED))
Pavel Emelyanovb4888932007-10-18 23:40:14 -07002010 uval = cmpxchg_futex_value_locked(uaddr, task_pid_vnr(current), 0);
Thomas Gleixner36cf3b52007-07-15 23:41:20 -07002011
Ingo Molnarc87e2832006-06-27 02:54:58 -07002012
2013 if (unlikely(uval == -EFAULT))
2014 goto pi_faulted;
2015 /*
2016 * Rare case: we managed to release the lock atomically,
2017 * no need to wake anyone else up:
2018 */
Pavel Emelyanovb4888932007-10-18 23:40:14 -07002019 if (unlikely(uval == task_pid_vnr(current)))
Ingo Molnarc87e2832006-06-27 02:54:58 -07002020 goto out_unlock;
2021
2022 /*
2023 * Ok, other tasks may need to be woken up - check waiters
2024 * and do the wakeup if necessary:
2025 */
2026 head = &hb->chain;
2027
Pierre Peifferec92d082007-05-09 02:35:00 -07002028 plist_for_each_entry_safe(this, next, head, list) {
Ingo Molnarc87e2832006-06-27 02:54:58 -07002029 if (!match_futex (&this->key, &key))
2030 continue;
2031 ret = wake_futex_pi(uaddr, uval, this);
2032 /*
2033 * The atomic access to the futex value
2034 * generated a pagefault, so retry the
2035 * user-access and the wakeup:
2036 */
2037 if (ret == -EFAULT)
2038 goto pi_faulted;
2039 goto out_unlock;
2040 }
2041 /*
2042 * No waiters - kernel unlocks the futex:
2043 */
Ingo Molnare3f2dde2006-07-29 05:17:57 +02002044 if (!(uval & FUTEX_OWNER_DIED)) {
2045 ret = unlock_futex_pi(uaddr, uval);
2046 if (ret == -EFAULT)
2047 goto pi_faulted;
2048 }
Ingo Molnarc87e2832006-06-27 02:54:58 -07002049
2050out_unlock:
2051 spin_unlock(&hb->lock);
Peter Zijlstra38d47c12008-09-26 19:32:20 +02002052 put_futex_key(fshared, &key);
Ingo Molnarc87e2832006-06-27 02:54:58 -07002053
Darren Hart42d35d42008-12-29 15:49:53 -08002054out:
Ingo Molnarc87e2832006-06-27 02:54:58 -07002055 return ret;
2056
2057pi_faulted:
Alexey Kuznetsov778e9a92007-06-08 13:47:00 -07002058 spin_unlock(&hb->lock);
Darren Harte4dc5b72009-03-12 00:56:13 -07002059 put_futex_key(fshared, &key);
Ingo Molnarc87e2832006-06-27 02:54:58 -07002060
Thomas Gleixnerd0725992009-06-11 23:15:43 +02002061 ret = fault_in_user_writeable(uaddr);
Darren Hartb5686362008-12-18 15:06:34 -08002062 if (!ret)
Ingo Molnarc87e2832006-06-27 02:54:58 -07002063 goto retry;
2064
Linus Torvalds1da177e2005-04-16 15:20:36 -07002065 return ret;
2066}
2067
Darren Hart52400ba2009-04-03 13:40:49 -07002068/**
2069 * handle_early_requeue_pi_wakeup() - Detect early wakeup on the initial futex
2070 * @hb: the hash_bucket futex_q was original enqueued on
2071 * @q: the futex_q woken while waiting to be requeued
2072 * @key2: the futex_key of the requeue target futex
2073 * @timeout: the timeout associated with the wait (NULL if none)
2074 *
2075 * Detect if the task was woken on the initial futex as opposed to the requeue
2076 * target futex. If so, determine if it was a timeout or a signal that caused
2077 * the wakeup and return the appropriate error code to the caller. Must be
2078 * called with the hb lock held.
2079 *
2080 * Returns
2081 * 0 - no early wakeup detected
Thomas Gleixner1c840c12009-05-20 09:22:40 +02002082 * <0 - -ETIMEDOUT or -ERESTARTNOINTR
Darren Hart52400ba2009-04-03 13:40:49 -07002083 */
2084static inline
2085int handle_early_requeue_pi_wakeup(struct futex_hash_bucket *hb,
2086 struct futex_q *q, union futex_key *key2,
2087 struct hrtimer_sleeper *timeout)
2088{
2089 int ret = 0;
2090
2091 /*
2092 * With the hb lock held, we avoid races while we process the wakeup.
2093 * We only need to hold hb (and not hb2) to ensure atomicity as the
2094 * wakeup code can't change q.key from uaddr to uaddr2 if we hold hb.
2095 * It can't be requeued from uaddr2 to something else since we don't
2096 * support a PI aware source futex for requeue.
2097 */
2098 if (!match_futex(&q->key, key2)) {
2099 WARN_ON(q->lock_ptr && (&hb->lock != q->lock_ptr));
2100 /*
2101 * We were woken prior to requeue by a timeout or a signal.
2102 * Unqueue the futex_q and determine which it was.
2103 */
2104 plist_del(&q->list, &q->list.plist);
2105 drop_futex_key_refs(&q->key);
2106
2107 if (timeout && !timeout->task)
2108 ret = -ETIMEDOUT;
Thomas Gleixner1c840c12009-05-20 09:22:40 +02002109 else
2110 ret = -ERESTARTNOINTR;
Darren Hart52400ba2009-04-03 13:40:49 -07002111 }
2112 return ret;
2113}
2114
2115/**
2116 * futex_wait_requeue_pi() - Wait on uaddr and take uaddr2
2117 * @uaddr: the futex we initialyl wait on (non-pi)
2118 * @fshared: whether the futexes are shared (1) or not (0). They must be
2119 * the same type, no requeueing from private to shared, etc.
2120 * @val: the expected value of uaddr
2121 * @abs_time: absolute timeout
2122 * @bitset: 32 bit wakeup bitset set by userspace, defaults to all.
2123 * @clockrt: whether to use CLOCK_REALTIME (1) or CLOCK_MONOTONIC (0)
2124 * @uaddr2: the pi futex we will take prior to returning to user-space
2125 *
2126 * The caller will wait on uaddr and will be requeued by futex_requeue() to
2127 * uaddr2 which must be PI aware. Normal wakeup will wake on uaddr2 and
2128 * complete the acquisition of the rt_mutex prior to returning to userspace.
2129 * This ensures the rt_mutex maintains an owner when it has waiters; without
2130 * one, the pi logic wouldn't know which task to boost/deboost, if there was a
2131 * need to.
2132 *
2133 * We call schedule in futex_wait_queue_me() when we enqueue and return there
2134 * via the following:
2135 * 1) wakeup on uaddr2 after an atomic lock acquisition by futex_requeue()
Darren Hartcc6db4e2009-07-31 16:20:10 -07002136 * 2) wakeup on uaddr2 after a requeue
2137 * 3) signal
2138 * 4) timeout
Darren Hart52400ba2009-04-03 13:40:49 -07002139 *
Darren Hartcc6db4e2009-07-31 16:20:10 -07002140 * If 3, cleanup and return -ERESTARTNOINTR.
Darren Hart52400ba2009-04-03 13:40:49 -07002141 *
2142 * If 2, we may then block on trying to take the rt_mutex and return via:
2143 * 5) successful lock
2144 * 6) signal
2145 * 7) timeout
2146 * 8) other lock acquisition failure
2147 *
Darren Hartcc6db4e2009-07-31 16:20:10 -07002148 * If 6, return -EWOULDBLOCK (restarting the syscall would do the same).
Darren Hart52400ba2009-04-03 13:40:49 -07002149 *
2150 * If 4 or 7, we cleanup and return with -ETIMEDOUT.
2151 *
2152 * Returns:
2153 * 0 - On success
2154 * <0 - On error
2155 */
2156static int futex_wait_requeue_pi(u32 __user *uaddr, int fshared,
2157 u32 val, ktime_t *abs_time, u32 bitset,
2158 int clockrt, u32 __user *uaddr2)
2159{
2160 struct hrtimer_sleeper timeout, *to = NULL;
2161 struct rt_mutex_waiter rt_waiter;
2162 struct rt_mutex *pi_mutex = NULL;
Darren Hart52400ba2009-04-03 13:40:49 -07002163 struct futex_hash_bucket *hb;
2164 union futex_key key2;
2165 struct futex_q q;
2166 int res, ret;
Darren Hart52400ba2009-04-03 13:40:49 -07002167
2168 if (!bitset)
2169 return -EINVAL;
2170
2171 if (abs_time) {
2172 to = &timeout;
2173 hrtimer_init_on_stack(&to->timer, clockrt ? CLOCK_REALTIME :
2174 CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
2175 hrtimer_init_sleeper(to, current);
2176 hrtimer_set_expires_range_ns(&to->timer, *abs_time,
2177 current->timer_slack_ns);
2178 }
2179
2180 /*
2181 * The waiter is allocated on our stack, manipulated by the requeue
2182 * code while we sleep on uaddr.
2183 */
2184 debug_rt_mutex_init_waiter(&rt_waiter);
2185 rt_waiter.task = NULL;
2186
Darren Hart52400ba2009-04-03 13:40:49 -07002187 key2 = FUTEX_KEY_INIT;
Thomas Gleixner521c1802009-05-20 09:02:28 +02002188 ret = get_futex_key(uaddr2, fshared, &key2, VERIFY_WRITE);
Darren Hart52400ba2009-04-03 13:40:49 -07002189 if (unlikely(ret != 0))
2190 goto out;
2191
Darren Hart84bc4af2009-08-13 17:36:53 -07002192 q.pi_state = NULL;
2193 q.bitset = bitset;
2194 q.rt_waiter = &rt_waiter;
2195 q.requeue_pi_key = &key2;
2196
Darren Hart52400ba2009-04-03 13:40:49 -07002197 /* Prepare to wait on uaddr. */
2198 ret = futex_wait_setup(uaddr, val, fshared, &q, &hb);
Thomas Gleixnerc8b15a72009-05-20 09:18:50 +02002199 if (ret)
2200 goto out_key2;
Darren Hart52400ba2009-04-03 13:40:49 -07002201
2202 /* Queue the futex_q, drop the hb lock, wait for wakeup. */
Thomas Gleixnerf1a11e02009-05-05 19:21:40 +02002203 futex_wait_queue_me(hb, &q, to);
Darren Hart52400ba2009-04-03 13:40:49 -07002204
2205 spin_lock(&hb->lock);
2206 ret = handle_early_requeue_pi_wakeup(hb, &q, &key2, to);
2207 spin_unlock(&hb->lock);
2208 if (ret)
2209 goto out_put_keys;
2210
2211 /*
2212 * In order for us to be here, we know our q.key == key2, and since
2213 * we took the hb->lock above, we also know that futex_requeue() has
2214 * completed and we no longer have to concern ourselves with a wakeup
2215 * race with the atomic proxy lock acquition by the requeue code.
2216 */
2217
2218 /* Check if the requeue code acquired the second futex for us. */
2219 if (!q.rt_waiter) {
2220 /*
2221 * Got the lock. We might not be the anticipated owner if we
2222 * did a lock-steal - fix up the PI-state in that case.
2223 */
2224 if (q.pi_state && (q.pi_state->owner != current)) {
2225 spin_lock(q.lock_ptr);
2226 ret = fixup_pi_state_owner(uaddr2, &q, current,
2227 fshared);
2228 spin_unlock(q.lock_ptr);
2229 }
2230 } else {
2231 /*
2232 * We have been woken up by futex_unlock_pi(), a timeout, or a
2233 * signal. futex_unlock_pi() will not destroy the lock_ptr nor
2234 * the pi_state.
2235 */
2236 WARN_ON(!&q.pi_state);
2237 pi_mutex = &q.pi_state->pi_mutex;
2238 ret = rt_mutex_finish_proxy_lock(pi_mutex, to, &rt_waiter, 1);
2239 debug_rt_mutex_free_waiter(&rt_waiter);
2240
2241 spin_lock(q.lock_ptr);
2242 /*
2243 * Fixup the pi_state owner and possibly acquire the lock if we
2244 * haven't already.
2245 */
2246 res = fixup_owner(uaddr2, fshared, &q, !ret);
2247 /*
2248 * If fixup_owner() returned an error, proprogate that. If it
2249 * acquired the lock, clear our -ETIMEDOUT or -EINTR.
2250 */
2251 if (res)
2252 ret = (res < 0) ? res : 0;
2253
2254 /* Unqueue and drop the lock. */
2255 unqueue_me_pi(&q);
2256 }
2257
2258 /*
2259 * If fixup_pi_state_owner() faulted and was unable to handle the
2260 * fault, unlock the rt_mutex and return the fault to userspace.
2261 */
2262 if (ret == -EFAULT) {
2263 if (rt_mutex_owner(pi_mutex) == current)
2264 rt_mutex_unlock(pi_mutex);
2265 } else if (ret == -EINTR) {
Darren Hart52400ba2009-04-03 13:40:49 -07002266 /*
Darren Hartcc6db4e2009-07-31 16:20:10 -07002267 * We've already been requeued, but cannot restart by calling
2268 * futex_lock_pi() directly. We could restart this syscall, but
2269 * it would detect that the user space "val" changed and return
2270 * -EWOULDBLOCK. Save the overhead of the restart and return
2271 * -EWOULDBLOCK directly.
Darren Hart52400ba2009-04-03 13:40:49 -07002272 */
Thomas Gleixner20708872009-05-19 23:04:59 +02002273 ret = -EWOULDBLOCK;
Darren Hart52400ba2009-04-03 13:40:49 -07002274 }
2275
2276out_put_keys:
2277 put_futex_key(fshared, &q.key);
Thomas Gleixnerc8b15a72009-05-20 09:18:50 +02002278out_key2:
Darren Hart52400ba2009-04-03 13:40:49 -07002279 put_futex_key(fshared, &key2);
2280
2281out:
2282 if (to) {
2283 hrtimer_cancel(&to->timer);
2284 destroy_hrtimer_on_stack(&to->timer);
2285 }
2286 return ret;
2287}
2288
Ingo Molnar0771dfe2006-03-27 01:16:22 -08002289/*
2290 * Support for robust futexes: the kernel cleans up held futexes at
2291 * thread exit time.
2292 *
2293 * Implementation: user-space maintains a per-thread list of locks it
2294 * is holding. Upon do_exit(), the kernel carefully walks this list,
2295 * and marks all locks that are owned by this thread with the
Ingo Molnarc87e2832006-06-27 02:54:58 -07002296 * FUTEX_OWNER_DIED bit, and wakes up a waiter (if any). The list is
Ingo Molnar0771dfe2006-03-27 01:16:22 -08002297 * always manipulated with the lock held, so the list is private and
2298 * per-thread. Userspace also maintains a per-thread 'list_op_pending'
2299 * field, to allow the kernel to clean up if the thread dies after
2300 * acquiring the lock, but just before it could have added itself to
2301 * the list. There can only be one such pending lock.
2302 */
2303
2304/**
2305 * sys_set_robust_list - set the robust-futex list head of a task
2306 * @head: pointer to the list-head
2307 * @len: length of the list-head, as userspace expects
2308 */
Heiko Carstens836f92a2009-01-14 14:14:33 +01002309SYSCALL_DEFINE2(set_robust_list, struct robust_list_head __user *, head,
2310 size_t, len)
Ingo Molnar0771dfe2006-03-27 01:16:22 -08002311{
Thomas Gleixnera0c1e902008-02-23 15:23:57 -08002312 if (!futex_cmpxchg_enabled)
2313 return -ENOSYS;
Ingo Molnar0771dfe2006-03-27 01:16:22 -08002314 /*
2315 * The kernel knows only one size for now:
2316 */
2317 if (unlikely(len != sizeof(*head)))
2318 return -EINVAL;
2319
2320 current->robust_list = head;
2321
2322 return 0;
2323}
2324
2325/**
2326 * sys_get_robust_list - get the robust-futex list head of a task
2327 * @pid: pid of the process [zero for current task]
2328 * @head_ptr: pointer to a list-head pointer, the kernel fills it in
2329 * @len_ptr: pointer to a length field, the kernel fills in the header size
2330 */
Heiko Carstens836f92a2009-01-14 14:14:33 +01002331SYSCALL_DEFINE3(get_robust_list, int, pid,
2332 struct robust_list_head __user * __user *, head_ptr,
2333 size_t __user *, len_ptr)
Ingo Molnar0771dfe2006-03-27 01:16:22 -08002334{
Al Viroba46df92006-10-10 22:46:07 +01002335 struct robust_list_head __user *head;
Ingo Molnar0771dfe2006-03-27 01:16:22 -08002336 unsigned long ret;
David Howellsc69e8d92008-11-14 10:39:19 +11002337 const struct cred *cred = current_cred(), *pcred;
Ingo Molnar0771dfe2006-03-27 01:16:22 -08002338
Thomas Gleixnera0c1e902008-02-23 15:23:57 -08002339 if (!futex_cmpxchg_enabled)
2340 return -ENOSYS;
2341
Ingo Molnar0771dfe2006-03-27 01:16:22 -08002342 if (!pid)
2343 head = current->robust_list;
2344 else {
2345 struct task_struct *p;
2346
2347 ret = -ESRCH;
Oleg Nesterovaaa2a972006-09-29 02:00:55 -07002348 rcu_read_lock();
Pavel Emelyanov228ebcb2007-10-18 23:40:16 -07002349 p = find_task_by_vpid(pid);
Ingo Molnar0771dfe2006-03-27 01:16:22 -08002350 if (!p)
2351 goto err_unlock;
2352 ret = -EPERM;
David Howellsc69e8d92008-11-14 10:39:19 +11002353 pcred = __task_cred(p);
2354 if (cred->euid != pcred->euid &&
2355 cred->euid != pcred->uid &&
David Howells76aac0e2008-11-14 10:39:12 +11002356 !capable(CAP_SYS_PTRACE))
Ingo Molnar0771dfe2006-03-27 01:16:22 -08002357 goto err_unlock;
2358 head = p->robust_list;
Oleg Nesterovaaa2a972006-09-29 02:00:55 -07002359 rcu_read_unlock();
Ingo Molnar0771dfe2006-03-27 01:16:22 -08002360 }
2361
2362 if (put_user(sizeof(*head), len_ptr))
2363 return -EFAULT;
2364 return put_user(head, head_ptr);
2365
2366err_unlock:
Oleg Nesterovaaa2a972006-09-29 02:00:55 -07002367 rcu_read_unlock();
Ingo Molnar0771dfe2006-03-27 01:16:22 -08002368
2369 return ret;
2370}
2371
2372/*
2373 * Process a futex-list entry, check whether it's owned by the
2374 * dying task, and do notification if so:
2375 */
Ingo Molnare3f2dde2006-07-29 05:17:57 +02002376int handle_futex_death(u32 __user *uaddr, struct task_struct *curr, int pi)
Ingo Molnar0771dfe2006-03-27 01:16:22 -08002377{
Ingo Molnare3f2dde2006-07-29 05:17:57 +02002378 u32 uval, nval, mval;
Ingo Molnar0771dfe2006-03-27 01:16:22 -08002379
Ingo Molnar8f17d3a2006-03-27 01:16:27 -08002380retry:
2381 if (get_user(uval, uaddr))
Ingo Molnar0771dfe2006-03-27 01:16:22 -08002382 return -1;
2383
Pavel Emelyanovb4888932007-10-18 23:40:14 -07002384 if ((uval & FUTEX_TID_MASK) == task_pid_vnr(curr)) {
Ingo Molnar0771dfe2006-03-27 01:16:22 -08002385 /*
2386 * Ok, this dying thread is truly holding a futex
2387 * of interest. Set the OWNER_DIED bit atomically
2388 * via cmpxchg, and if the value had FUTEX_WAITERS
2389 * set, wake up a waiter (if any). (We have to do a
2390 * futex_wake() even if OWNER_DIED is already set -
2391 * to handle the rare but possible case of recursive
2392 * thread-death.) The rest of the cleanup is done in
2393 * userspace.
2394 */
Ingo Molnare3f2dde2006-07-29 05:17:57 +02002395 mval = (uval & FUTEX_WAITERS) | FUTEX_OWNER_DIED;
2396 nval = futex_atomic_cmpxchg_inatomic(uaddr, uval, mval);
2397
Ingo Molnarc87e2832006-06-27 02:54:58 -07002398 if (nval == -EFAULT)
2399 return -1;
2400
2401 if (nval != uval)
Ingo Molnar8f17d3a2006-03-27 01:16:27 -08002402 goto retry;
Ingo Molnar0771dfe2006-03-27 01:16:22 -08002403
Ingo Molnare3f2dde2006-07-29 05:17:57 +02002404 /*
2405 * Wake robust non-PI futexes here. The wakeup of
2406 * PI futexes happens in exit_pi_state():
2407 */
Thomas Gleixner36cf3b52007-07-15 23:41:20 -07002408 if (!pi && (uval & FUTEX_WAITERS))
Peter Zijlstrac2f9f202008-09-26 19:32:23 +02002409 futex_wake(uaddr, 1, 1, FUTEX_BITSET_MATCH_ANY);
Ingo Molnar0771dfe2006-03-27 01:16:22 -08002410 }
2411 return 0;
2412}
2413
2414/*
Ingo Molnare3f2dde2006-07-29 05:17:57 +02002415 * Fetch a robust-list pointer. Bit 0 signals PI futexes:
2416 */
2417static inline int fetch_robust_entry(struct robust_list __user **entry,
Al Viroba46df92006-10-10 22:46:07 +01002418 struct robust_list __user * __user *head,
2419 int *pi)
Ingo Molnare3f2dde2006-07-29 05:17:57 +02002420{
2421 unsigned long uentry;
2422
Al Viroba46df92006-10-10 22:46:07 +01002423 if (get_user(uentry, (unsigned long __user *)head))
Ingo Molnare3f2dde2006-07-29 05:17:57 +02002424 return -EFAULT;
2425
Al Viroba46df92006-10-10 22:46:07 +01002426 *entry = (void __user *)(uentry & ~1UL);
Ingo Molnare3f2dde2006-07-29 05:17:57 +02002427 *pi = uentry & 1;
2428
2429 return 0;
2430}
2431
2432/*
Ingo Molnar0771dfe2006-03-27 01:16:22 -08002433 * Walk curr->robust_list (very carefully, it's a userspace list!)
2434 * and mark any locks found there dead, and notify any waiters.
2435 *
2436 * We silently return on any sign of list-walking problem.
2437 */
2438void exit_robust_list(struct task_struct *curr)
2439{
2440 struct robust_list_head __user *head = curr->robust_list;
Martin Schwidefsky9f96cb12007-10-01 01:20:13 -07002441 struct robust_list __user *entry, *next_entry, *pending;
2442 unsigned int limit = ROBUST_LIST_LIMIT, pi, next_pi, pip;
Ingo Molnar0771dfe2006-03-27 01:16:22 -08002443 unsigned long futex_offset;
Martin Schwidefsky9f96cb12007-10-01 01:20:13 -07002444 int rc;
Ingo Molnar0771dfe2006-03-27 01:16:22 -08002445
Thomas Gleixnera0c1e902008-02-23 15:23:57 -08002446 if (!futex_cmpxchg_enabled)
2447 return;
2448
Ingo Molnar0771dfe2006-03-27 01:16:22 -08002449 /*
2450 * Fetch the list head (which was registered earlier, via
2451 * sys_set_robust_list()):
2452 */
Ingo Molnare3f2dde2006-07-29 05:17:57 +02002453 if (fetch_robust_entry(&entry, &head->list.next, &pi))
Ingo Molnar0771dfe2006-03-27 01:16:22 -08002454 return;
2455 /*
2456 * Fetch the relative futex offset:
2457 */
2458 if (get_user(futex_offset, &head->futex_offset))
2459 return;
2460 /*
2461 * Fetch any possibly pending lock-add first, and handle it
2462 * if it exists:
2463 */
Ingo Molnare3f2dde2006-07-29 05:17:57 +02002464 if (fetch_robust_entry(&pending, &head->list_op_pending, &pip))
Ingo Molnar0771dfe2006-03-27 01:16:22 -08002465 return;
Ingo Molnare3f2dde2006-07-29 05:17:57 +02002466
Martin Schwidefsky9f96cb12007-10-01 01:20:13 -07002467 next_entry = NULL; /* avoid warning with gcc */
Ingo Molnar0771dfe2006-03-27 01:16:22 -08002468 while (entry != &head->list) {
2469 /*
Martin Schwidefsky9f96cb12007-10-01 01:20:13 -07002470 * Fetch the next entry in the list before calling
2471 * handle_futex_death:
2472 */
2473 rc = fetch_robust_entry(&next_entry, &entry->next, &next_pi);
2474 /*
Ingo Molnar0771dfe2006-03-27 01:16:22 -08002475 * A pending lock might already be on the list, so
Ingo Molnarc87e2832006-06-27 02:54:58 -07002476 * don't process it twice:
Ingo Molnar0771dfe2006-03-27 01:16:22 -08002477 */
2478 if (entry != pending)
Al Viroba46df92006-10-10 22:46:07 +01002479 if (handle_futex_death((void __user *)entry + futex_offset,
Ingo Molnare3f2dde2006-07-29 05:17:57 +02002480 curr, pi))
Ingo Molnar0771dfe2006-03-27 01:16:22 -08002481 return;
Martin Schwidefsky9f96cb12007-10-01 01:20:13 -07002482 if (rc)
Ingo Molnar0771dfe2006-03-27 01:16:22 -08002483 return;
Martin Schwidefsky9f96cb12007-10-01 01:20:13 -07002484 entry = next_entry;
2485 pi = next_pi;
Ingo Molnar0771dfe2006-03-27 01:16:22 -08002486 /*
2487 * Avoid excessively long or circular lists:
2488 */
2489 if (!--limit)
2490 break;
2491
2492 cond_resched();
2493 }
Martin Schwidefsky9f96cb12007-10-01 01:20:13 -07002494
2495 if (pending)
2496 handle_futex_death((void __user *)pending + futex_offset,
2497 curr, pip);
Ingo Molnar0771dfe2006-03-27 01:16:22 -08002498}
2499
Pierre Peifferc19384b2007-05-09 02:35:02 -07002500long do_futex(u32 __user *uaddr, int op, u32 val, ktime_t *timeout,
Ingo Molnare2970f22006-06-27 02:54:47 -07002501 u32 __user *uaddr2, u32 val2, u32 val3)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002502{
Thomas Gleixner1acdac12008-11-20 10:02:53 -08002503 int clockrt, ret = -ENOSYS;
Eric Dumazet34f01cc2007-05-09 02:35:04 -07002504 int cmd = op & FUTEX_CMD_MASK;
Peter Zijlstrac2f9f202008-09-26 19:32:23 +02002505 int fshared = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002506
Eric Dumazet34f01cc2007-05-09 02:35:04 -07002507 if (!(op & FUTEX_PRIVATE_FLAG))
Peter Zijlstrac2f9f202008-09-26 19:32:23 +02002508 fshared = 1;
Eric Dumazet34f01cc2007-05-09 02:35:04 -07002509
Thomas Gleixner1acdac12008-11-20 10:02:53 -08002510 clockrt = op & FUTEX_CLOCK_REALTIME;
Darren Hart52400ba2009-04-03 13:40:49 -07002511 if (clockrt && cmd != FUTEX_WAIT_BITSET && cmd != FUTEX_WAIT_REQUEUE_PI)
Thomas Gleixner1acdac12008-11-20 10:02:53 -08002512 return -ENOSYS;
Eric Dumazet34f01cc2007-05-09 02:35:04 -07002513
2514 switch (cmd) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002515 case FUTEX_WAIT:
Thomas Gleixnercd689982008-02-01 17:45:14 +01002516 val3 = FUTEX_BITSET_MATCH_ANY;
2517 case FUTEX_WAIT_BITSET:
Thomas Gleixner1acdac12008-11-20 10:02:53 -08002518 ret = futex_wait(uaddr, fshared, val, timeout, val3, clockrt);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002519 break;
2520 case FUTEX_WAKE:
Thomas Gleixnercd689982008-02-01 17:45:14 +01002521 val3 = FUTEX_BITSET_MATCH_ANY;
2522 case FUTEX_WAKE_BITSET:
2523 ret = futex_wake(uaddr, fshared, val, val3);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002524 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002525 case FUTEX_REQUEUE:
Darren Hart52400ba2009-04-03 13:40:49 -07002526 ret = futex_requeue(uaddr, fshared, uaddr2, val, val2, NULL, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002527 break;
2528 case FUTEX_CMP_REQUEUE:
Darren Hart52400ba2009-04-03 13:40:49 -07002529 ret = futex_requeue(uaddr, fshared, uaddr2, val, val2, &val3,
2530 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002531 break;
Jakub Jelinek4732efb2005-09-06 15:16:25 -07002532 case FUTEX_WAKE_OP:
Eric Dumazet34f01cc2007-05-09 02:35:04 -07002533 ret = futex_wake_op(uaddr, fshared, uaddr2, val, val2, val3);
Jakub Jelinek4732efb2005-09-06 15:16:25 -07002534 break;
Ingo Molnarc87e2832006-06-27 02:54:58 -07002535 case FUTEX_LOCK_PI:
Thomas Gleixnera0c1e902008-02-23 15:23:57 -08002536 if (futex_cmpxchg_enabled)
2537 ret = futex_lock_pi(uaddr, fshared, val, timeout, 0);
Ingo Molnarc87e2832006-06-27 02:54:58 -07002538 break;
2539 case FUTEX_UNLOCK_PI:
Thomas Gleixnera0c1e902008-02-23 15:23:57 -08002540 if (futex_cmpxchg_enabled)
2541 ret = futex_unlock_pi(uaddr, fshared);
Ingo Molnarc87e2832006-06-27 02:54:58 -07002542 break;
2543 case FUTEX_TRYLOCK_PI:
Thomas Gleixnera0c1e902008-02-23 15:23:57 -08002544 if (futex_cmpxchg_enabled)
2545 ret = futex_lock_pi(uaddr, fshared, 0, timeout, 1);
Ingo Molnarc87e2832006-06-27 02:54:58 -07002546 break;
Darren Hart52400ba2009-04-03 13:40:49 -07002547 case FUTEX_WAIT_REQUEUE_PI:
2548 val3 = FUTEX_BITSET_MATCH_ANY;
2549 ret = futex_wait_requeue_pi(uaddr, fshared, val, timeout, val3,
2550 clockrt, uaddr2);
2551 break;
Darren Hart52400ba2009-04-03 13:40:49 -07002552 case FUTEX_CMP_REQUEUE_PI:
2553 ret = futex_requeue(uaddr, fshared, uaddr2, val, val2, &val3,
2554 1);
2555 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002556 default:
2557 ret = -ENOSYS;
2558 }
2559 return ret;
2560}
2561
2562
Heiko Carstens17da2bd2009-01-14 14:14:10 +01002563SYSCALL_DEFINE6(futex, u32 __user *, uaddr, int, op, u32, val,
2564 struct timespec __user *, utime, u32 __user *, uaddr2,
2565 u32, val3)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002566{
Pierre Peifferc19384b2007-05-09 02:35:02 -07002567 struct timespec ts;
2568 ktime_t t, *tp = NULL;
Ingo Molnare2970f22006-06-27 02:54:47 -07002569 u32 val2 = 0;
Eric Dumazet34f01cc2007-05-09 02:35:04 -07002570 int cmd = op & FUTEX_CMD_MASK;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002571
Thomas Gleixnercd689982008-02-01 17:45:14 +01002572 if (utime && (cmd == FUTEX_WAIT || cmd == FUTEX_LOCK_PI ||
Darren Hart52400ba2009-04-03 13:40:49 -07002573 cmd == FUTEX_WAIT_BITSET ||
2574 cmd == FUTEX_WAIT_REQUEUE_PI)) {
Pierre Peifferc19384b2007-05-09 02:35:02 -07002575 if (copy_from_user(&ts, utime, sizeof(ts)) != 0)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002576 return -EFAULT;
Pierre Peifferc19384b2007-05-09 02:35:02 -07002577 if (!timespec_valid(&ts))
Thomas Gleixner9741ef962006-03-31 02:31:32 -08002578 return -EINVAL;
Pierre Peifferc19384b2007-05-09 02:35:02 -07002579
2580 t = timespec_to_ktime(ts);
Eric Dumazet34f01cc2007-05-09 02:35:04 -07002581 if (cmd == FUTEX_WAIT)
Thomas Gleixner5a7780e2008-02-13 09:20:43 +01002582 t = ktime_add_safe(ktime_get(), t);
Pierre Peifferc19384b2007-05-09 02:35:02 -07002583 tp = &t;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002584 }
2585 /*
Darren Hart52400ba2009-04-03 13:40:49 -07002586 * requeue parameter in 'utime' if cmd == FUTEX_*_REQUEUE_*.
Andreas Schwabf54f0982007-07-31 00:38:51 -07002587 * number of waiters to wake in 'utime' if cmd == FUTEX_WAKE_OP.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002588 */
Andreas Schwabf54f0982007-07-31 00:38:51 -07002589 if (cmd == FUTEX_REQUEUE || cmd == FUTEX_CMP_REQUEUE ||
Darren Hartba9c22f2009-04-20 22:22:22 -07002590 cmd == FUTEX_CMP_REQUEUE_PI || cmd == FUTEX_WAKE_OP)
Ingo Molnare2970f22006-06-27 02:54:47 -07002591 val2 = (u32) (unsigned long) utime;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002592
Pierre Peifferc19384b2007-05-09 02:35:02 -07002593 return do_futex(uaddr, op, val, tp, uaddr2, val2, val3);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002594}
2595
Benjamin Herrenschmidtf6d107f2008-03-27 14:52:15 +11002596static int __init futex_init(void)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002597{
Thomas Gleixnera0c1e902008-02-23 15:23:57 -08002598 u32 curval;
Thomas Gleixner3e4ab742008-02-23 15:23:55 -08002599 int i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002600
Thomas Gleixnera0c1e902008-02-23 15:23:57 -08002601 /*
2602 * This will fail and we want it. Some arch implementations do
2603 * runtime detection of the futex_atomic_cmpxchg_inatomic()
2604 * functionality. We want to know that before we call in any
2605 * of the complex code paths. Also we want to prevent
2606 * registration of robust lists in that case. NULL is
2607 * guaranteed to fault and we get -EFAULT on functional
2608 * implementation, the non functional ones will return
2609 * -ENOSYS.
2610 */
2611 curval = cmpxchg_futex_value_locked(NULL, 0, 0);
2612 if (curval == -EFAULT)
2613 futex_cmpxchg_enabled = 1;
2614
Thomas Gleixner3e4ab742008-02-23 15:23:55 -08002615 for (i = 0; i < ARRAY_SIZE(futex_queues); i++) {
2616 plist_head_init(&futex_queues[i].chain, &futex_queues[i].lock);
2617 spin_lock_init(&futex_queues[i].lock);
2618 }
2619
Linus Torvalds1da177e2005-04-16 15:20:36 -07002620 return 0;
2621}
Benjamin Herrenschmidtf6d107f2008-03-27 14:52:15 +11002622__initcall(futex_init);