xref: /linux/fs/afs/flock.c (revision 0dd9ac63ce26ec87b080ca9c3e6efed33c23ace6)
1 /* AFS file locking support
2  *
3  * Copyright (C) 2007 Red Hat, Inc. All Rights Reserved.
4  * Written by David Howells (dhowells@redhat.com)
5  *
6  * This program is free software; you can redistribute it and/or
7  * modify it under the terms of the GNU General Public License
8  * as published by the Free Software Foundation; either version
9  * 2 of the License, or (at your option) any later version.
10  */
11 
12 #include <linux/smp_lock.h>
13 #include "internal.h"
14 
15 #define AFS_LOCK_GRANTED	0
16 #define AFS_LOCK_PENDING	1
17 
18 static void afs_fl_copy_lock(struct file_lock *new, struct file_lock *fl);
19 static void afs_fl_release_private(struct file_lock *fl);
20 
21 static struct workqueue_struct *afs_lock_manager;
22 static DEFINE_MUTEX(afs_lock_manager_mutex);
23 
24 static const struct file_lock_operations afs_lock_ops = {
25 	.fl_copy_lock		= afs_fl_copy_lock,
26 	.fl_release_private	= afs_fl_release_private,
27 };
28 
29 /*
30  * initialise the lock manager thread if it isn't already running
31  */
32 static int afs_init_lock_manager(void)
33 {
34 	int ret;
35 
36 	ret = 0;
37 	if (!afs_lock_manager) {
38 		mutex_lock(&afs_lock_manager_mutex);
39 		if (!afs_lock_manager) {
40 			afs_lock_manager =
41 				create_singlethread_workqueue("kafs_lockd");
42 			if (!afs_lock_manager)
43 				ret = -ENOMEM;
44 		}
45 		mutex_unlock(&afs_lock_manager_mutex);
46 	}
47 	return ret;
48 }
49 
50 /*
51  * destroy the lock manager thread if it's running
52  */
53 void __exit afs_kill_lock_manager(void)
54 {
55 	if (afs_lock_manager)
56 		destroy_workqueue(afs_lock_manager);
57 }
58 
59 /*
60  * if the callback is broken on this vnode, then the lock may now be available
61  */
62 void afs_lock_may_be_available(struct afs_vnode *vnode)
63 {
64 	_enter("{%x:%u}", vnode->fid.vid, vnode->fid.vnode);
65 
66 	queue_delayed_work(afs_lock_manager, &vnode->lock_work, 0);
67 }
68 
69 /*
70  * the lock will time out in 5 minutes unless we extend it, so schedule
71  * extension in a bit less than that time
72  */
73 static void afs_schedule_lock_extension(struct afs_vnode *vnode)
74 {
75 	queue_delayed_work(afs_lock_manager, &vnode->lock_work,
76 			   AFS_LOCKWAIT * HZ / 2);
77 }
78 
79 /*
80  * grant one or more locks (readlocks are allowed to jump the queue if the
81  * first lock in the queue is itself a readlock)
82  * - the caller must hold the vnode lock
83  */
84 static void afs_grant_locks(struct afs_vnode *vnode, struct file_lock *fl)
85 {
86 	struct file_lock *p, *_p;
87 
88 	list_move_tail(&fl->fl_u.afs.link, &vnode->granted_locks);
89 	if (fl->fl_type == F_RDLCK) {
90 		list_for_each_entry_safe(p, _p, &vnode->pending_locks,
91 					 fl_u.afs.link) {
92 			if (p->fl_type == F_RDLCK) {
93 				p->fl_u.afs.state = AFS_LOCK_GRANTED;
94 				list_move_tail(&p->fl_u.afs.link,
95 					       &vnode->granted_locks);
96 				wake_up(&p->fl_wait);
97 			}
98 		}
99 	}
100 }
101 
102 /*
103  * do work for a lock, including:
104  * - probing for a lock we're waiting on but didn't get immediately
105  * - extending a lock that's close to timing out
106  */
107 void afs_lock_work(struct work_struct *work)
108 {
109 	struct afs_vnode *vnode =
110 		container_of(work, struct afs_vnode, lock_work.work);
111 	struct file_lock *fl;
112 	afs_lock_type_t type;
113 	struct key *key;
114 	int ret;
115 
116 	_enter("{%x:%u}", vnode->fid.vid, vnode->fid.vnode);
117 
118 	spin_lock(&vnode->lock);
119 
120 	if (test_bit(AFS_VNODE_UNLOCKING, &vnode->flags)) {
121 		_debug("unlock");
122 		spin_unlock(&vnode->lock);
123 
124 		/* attempt to release the server lock; if it fails, we just
125 		 * wait 5 minutes and it'll time out anyway */
126 		ret = afs_vnode_release_lock(vnode, vnode->unlock_key);
127 		if (ret < 0)
128 			printk(KERN_WARNING "AFS:"
129 			       " Failed to release lock on {%x:%x} error %d\n",
130 			       vnode->fid.vid, vnode->fid.vnode, ret);
131 
132 		spin_lock(&vnode->lock);
133 		key_put(vnode->unlock_key);
134 		vnode->unlock_key = NULL;
135 		clear_bit(AFS_VNODE_UNLOCKING, &vnode->flags);
136 	}
137 
138 	/* if we've got a lock, then it must be time to extend that lock as AFS
139 	 * locks time out after 5 minutes */
140 	if (!list_empty(&vnode->granted_locks)) {
141 		_debug("extend");
142 
143 		if (test_and_set_bit(AFS_VNODE_LOCKING, &vnode->flags))
144 			BUG();
145 		fl = list_entry(vnode->granted_locks.next,
146 				struct file_lock, fl_u.afs.link);
147 		key = key_get(fl->fl_file->private_data);
148 		spin_unlock(&vnode->lock);
149 
150 		ret = afs_vnode_extend_lock(vnode, key);
151 		clear_bit(AFS_VNODE_LOCKING, &vnode->flags);
152 		key_put(key);
153 		switch (ret) {
154 		case 0:
155 			afs_schedule_lock_extension(vnode);
156 			break;
157 		default:
158 			/* ummm... we failed to extend the lock - retry
159 			 * extension shortly */
160 			printk(KERN_WARNING "AFS:"
161 			       " Failed to extend lock on {%x:%x} error %d\n",
162 			       vnode->fid.vid, vnode->fid.vnode, ret);
163 			queue_delayed_work(afs_lock_manager, &vnode->lock_work,
164 					   HZ * 10);
165 			break;
166 		}
167 		_leave(" [extend]");
168 		return;
169 	}
170 
171 	/* if we don't have a granted lock, then we must've been called back by
172 	 * the server, and so if might be possible to get a lock we're
173 	 * currently waiting for */
174 	if (!list_empty(&vnode->pending_locks)) {
175 		_debug("get");
176 
177 		if (test_and_set_bit(AFS_VNODE_LOCKING, &vnode->flags))
178 			BUG();
179 		fl = list_entry(vnode->pending_locks.next,
180 				struct file_lock, fl_u.afs.link);
181 		key = key_get(fl->fl_file->private_data);
182 		type = (fl->fl_type == F_RDLCK) ?
183 			AFS_LOCK_READ : AFS_LOCK_WRITE;
184 		spin_unlock(&vnode->lock);
185 
186 		ret = afs_vnode_set_lock(vnode, key, type);
187 		clear_bit(AFS_VNODE_LOCKING, &vnode->flags);
188 		switch (ret) {
189 		case -EWOULDBLOCK:
190 			_debug("blocked");
191 			break;
192 		case 0:
193 			_debug("acquired");
194 			if (type == AFS_LOCK_READ)
195 				set_bit(AFS_VNODE_READLOCKED, &vnode->flags);
196 			else
197 				set_bit(AFS_VNODE_WRITELOCKED, &vnode->flags);
198 			ret = AFS_LOCK_GRANTED;
199 		default:
200 			spin_lock(&vnode->lock);
201 			/* the pending lock may have been withdrawn due to a
202 			 * signal */
203 			if (list_entry(vnode->pending_locks.next,
204 				       struct file_lock, fl_u.afs.link) == fl) {
205 				fl->fl_u.afs.state = ret;
206 				if (ret == AFS_LOCK_GRANTED)
207 					afs_grant_locks(vnode, fl);
208 				else
209 					list_del_init(&fl->fl_u.afs.link);
210 				wake_up(&fl->fl_wait);
211 				spin_unlock(&vnode->lock);
212 			} else {
213 				_debug("withdrawn");
214 				clear_bit(AFS_VNODE_READLOCKED, &vnode->flags);
215 				clear_bit(AFS_VNODE_WRITELOCKED, &vnode->flags);
216 				spin_unlock(&vnode->lock);
217 				afs_vnode_release_lock(vnode, key);
218 				if (!list_empty(&vnode->pending_locks))
219 					afs_lock_may_be_available(vnode);
220 			}
221 			break;
222 		}
223 		key_put(key);
224 		_leave(" [pend]");
225 		return;
226 	}
227 
228 	/* looks like the lock request was withdrawn on a signal */
229 	spin_unlock(&vnode->lock);
230 	_leave(" [no locks]");
231 }
232 
233 /*
234  * pass responsibility for the unlocking of a vnode on the server to the
235  * manager thread, lest a pending signal in the calling thread interrupt
236  * AF_RXRPC
237  * - the caller must hold the vnode lock
238  */
239 static void afs_defer_unlock(struct afs_vnode *vnode, struct key *key)
240 {
241 	cancel_delayed_work(&vnode->lock_work);
242 	if (!test_and_clear_bit(AFS_VNODE_READLOCKED, &vnode->flags) &&
243 	    !test_and_clear_bit(AFS_VNODE_WRITELOCKED, &vnode->flags))
244 		BUG();
245 	if (test_and_set_bit(AFS_VNODE_UNLOCKING, &vnode->flags))
246 		BUG();
247 	vnode->unlock_key = key_get(key);
248 	afs_lock_may_be_available(vnode);
249 }
250 
251 /*
252  * request a lock on a file on the server
253  */
254 static int afs_do_setlk(struct file *file, struct file_lock *fl)
255 {
256 	struct afs_vnode *vnode = AFS_FS_I(file->f_mapping->host);
257 	afs_lock_type_t type;
258 	struct key *key = file->private_data;
259 	int ret;
260 
261 	_enter("{%x:%u},%u", vnode->fid.vid, vnode->fid.vnode, fl->fl_type);
262 
263 	/* only whole-file locks are supported */
264 	if (fl->fl_start != 0 || fl->fl_end != OFFSET_MAX)
265 		return -EINVAL;
266 
267 	ret = afs_init_lock_manager();
268 	if (ret < 0)
269 		return ret;
270 
271 	fl->fl_ops = &afs_lock_ops;
272 	INIT_LIST_HEAD(&fl->fl_u.afs.link);
273 	fl->fl_u.afs.state = AFS_LOCK_PENDING;
274 
275 	type = (fl->fl_type == F_RDLCK) ? AFS_LOCK_READ : AFS_LOCK_WRITE;
276 
277 	lock_kernel();
278 
279 	/* make sure we've got a callback on this file and that our view of the
280 	 * data version is up to date */
281 	ret = afs_vnode_fetch_status(vnode, NULL, key);
282 	if (ret < 0)
283 		goto error;
284 
285 	if (vnode->status.lock_count != 0 && !(fl->fl_flags & FL_SLEEP)) {
286 		ret = -EAGAIN;
287 		goto error;
288 	}
289 
290 	spin_lock(&vnode->lock);
291 
292 	/* if we've already got a readlock on the server then we can instantly
293 	 * grant another readlock, irrespective of whether there are any
294 	 * pending writelocks */
295 	if (type == AFS_LOCK_READ &&
296 	    vnode->flags & (1 << AFS_VNODE_READLOCKED)) {
297 		_debug("instant readlock");
298 		ASSERTCMP(vnode->flags &
299 			  ((1 << AFS_VNODE_LOCKING) |
300 			   (1 << AFS_VNODE_WRITELOCKED)), ==, 0);
301 		ASSERT(!list_empty(&vnode->granted_locks));
302 		goto sharing_existing_lock;
303 	}
304 
305 	/* if there's no-one else with a lock on this vnode, then we need to
306 	 * ask the server for a lock */
307 	if (list_empty(&vnode->pending_locks) &&
308 	    list_empty(&vnode->granted_locks)) {
309 		_debug("not locked");
310 		ASSERTCMP(vnode->flags &
311 			  ((1 << AFS_VNODE_LOCKING) |
312 			   (1 << AFS_VNODE_READLOCKED) |
313 			   (1 << AFS_VNODE_WRITELOCKED)), ==, 0);
314 		list_add_tail(&fl->fl_u.afs.link, &vnode->pending_locks);
315 		set_bit(AFS_VNODE_LOCKING, &vnode->flags);
316 		spin_unlock(&vnode->lock);
317 
318 		ret = afs_vnode_set_lock(vnode, key, type);
319 		clear_bit(AFS_VNODE_LOCKING, &vnode->flags);
320 		switch (ret) {
321 		case 0:
322 			_debug("acquired");
323 			goto acquired_server_lock;
324 		case -EWOULDBLOCK:
325 			_debug("would block");
326 			spin_lock(&vnode->lock);
327 			ASSERT(list_empty(&vnode->granted_locks));
328 			ASSERTCMP(vnode->pending_locks.next, ==,
329 				  &fl->fl_u.afs.link);
330 			goto wait;
331 		default:
332 			spin_lock(&vnode->lock);
333 			list_del_init(&fl->fl_u.afs.link);
334 			spin_unlock(&vnode->lock);
335 			goto error;
336 		}
337 	}
338 
339 	/* otherwise, we need to wait for a local lock to become available */
340 	_debug("wait local");
341 	list_add_tail(&fl->fl_u.afs.link, &vnode->pending_locks);
342 wait:
343 	if (!(fl->fl_flags & FL_SLEEP)) {
344 		_debug("noblock");
345 		ret = -EAGAIN;
346 		goto abort_attempt;
347 	}
348 	spin_unlock(&vnode->lock);
349 
350 	/* now we need to sleep and wait for the lock manager thread to get the
351 	 * lock from the server */
352 	_debug("sleep");
353 	ret = wait_event_interruptible(fl->fl_wait,
354 				       fl->fl_u.afs.state <= AFS_LOCK_GRANTED);
355 	if (fl->fl_u.afs.state <= AFS_LOCK_GRANTED) {
356 		ret = fl->fl_u.afs.state;
357 		if (ret < 0)
358 			goto error;
359 		spin_lock(&vnode->lock);
360 		goto given_lock;
361 	}
362 
363 	/* we were interrupted, but someone may still be in the throes of
364 	 * giving us the lock */
365 	_debug("intr");
366 	ASSERTCMP(ret, ==, -ERESTARTSYS);
367 
368 	spin_lock(&vnode->lock);
369 	if (fl->fl_u.afs.state <= AFS_LOCK_GRANTED) {
370 		ret = fl->fl_u.afs.state;
371 		if (ret < 0) {
372 			spin_unlock(&vnode->lock);
373 			goto error;
374 		}
375 		goto given_lock;
376 	}
377 
378 abort_attempt:
379 	/* we aren't going to get the lock, either because we're unwilling to
380 	 * wait, or because some signal happened */
381 	_debug("abort");
382 	if (list_empty(&vnode->granted_locks) &&
383 	    vnode->pending_locks.next == &fl->fl_u.afs.link) {
384 		if (vnode->pending_locks.prev != &fl->fl_u.afs.link) {
385 			/* kick the next pending lock into having a go */
386 			list_del_init(&fl->fl_u.afs.link);
387 			afs_lock_may_be_available(vnode);
388 		}
389 	} else {
390 		list_del_init(&fl->fl_u.afs.link);
391 	}
392 	spin_unlock(&vnode->lock);
393 	goto error;
394 
395 acquired_server_lock:
396 	/* we've acquired a server lock, but it needs to be renewed after 5
397 	 * mins */
398 	spin_lock(&vnode->lock);
399 	afs_schedule_lock_extension(vnode);
400 	if (type == AFS_LOCK_READ)
401 		set_bit(AFS_VNODE_READLOCKED, &vnode->flags);
402 	else
403 		set_bit(AFS_VNODE_WRITELOCKED, &vnode->flags);
404 sharing_existing_lock:
405 	/* the lock has been granted as far as we're concerned... */
406 	fl->fl_u.afs.state = AFS_LOCK_GRANTED;
407 	list_move_tail(&fl->fl_u.afs.link, &vnode->granted_locks);
408 given_lock:
409 	/* ... but we do still need to get the VFS's blessing */
410 	ASSERT(!(vnode->flags & (1 << AFS_VNODE_LOCKING)));
411 	ASSERT((vnode->flags & ((1 << AFS_VNODE_READLOCKED) |
412 				(1 << AFS_VNODE_WRITELOCKED))) != 0);
413 	ret = posix_lock_file(file, fl, NULL);
414 	if (ret < 0)
415 		goto vfs_rejected_lock;
416 	spin_unlock(&vnode->lock);
417 
418 	/* again, make sure we've got a callback on this file and, again, make
419 	 * sure that our view of the data version is up to date (we ignore
420 	 * errors incurred here and deal with the consequences elsewhere) */
421 	afs_vnode_fetch_status(vnode, NULL, key);
422 
423 error:
424 	unlock_kernel();
425 	_leave(" = %d", ret);
426 	return ret;
427 
428 vfs_rejected_lock:
429 	/* the VFS rejected the lock we just obtained, so we have to discard
430 	 * what we just got */
431 	_debug("vfs refused %d", ret);
432 	list_del_init(&fl->fl_u.afs.link);
433 	if (list_empty(&vnode->granted_locks))
434 		afs_defer_unlock(vnode, key);
435 	goto abort_attempt;
436 }
437 
438 /*
439  * unlock on a file on the server
440  */
441 static int afs_do_unlk(struct file *file, struct file_lock *fl)
442 {
443 	struct afs_vnode *vnode = AFS_FS_I(file->f_mapping->host);
444 	struct key *key = file->private_data;
445 	int ret;
446 
447 	_enter("{%x:%u},%u", vnode->fid.vid, vnode->fid.vnode, fl->fl_type);
448 
449 	/* only whole-file unlocks are supported */
450 	if (fl->fl_start != 0 || fl->fl_end != OFFSET_MAX)
451 		return -EINVAL;
452 
453 	fl->fl_ops = &afs_lock_ops;
454 	INIT_LIST_HEAD(&fl->fl_u.afs.link);
455 	fl->fl_u.afs.state = AFS_LOCK_PENDING;
456 
457 	spin_lock(&vnode->lock);
458 	ret = posix_lock_file(file, fl, NULL);
459 	if (ret < 0) {
460 		spin_unlock(&vnode->lock);
461 		_leave(" = %d [vfs]", ret);
462 		return ret;
463 	}
464 
465 	/* discard the server lock only if all granted locks are gone */
466 	if (list_empty(&vnode->granted_locks))
467 		afs_defer_unlock(vnode, key);
468 	spin_unlock(&vnode->lock);
469 	_leave(" = 0");
470 	return 0;
471 }
472 
473 /*
474  * return information about a lock we currently hold, if indeed we hold one
475  */
476 static int afs_do_getlk(struct file *file, struct file_lock *fl)
477 {
478 	struct afs_vnode *vnode = AFS_FS_I(file->f_mapping->host);
479 	struct key *key = file->private_data;
480 	int ret, lock_count;
481 
482 	_enter("");
483 
484 	fl->fl_type = F_UNLCK;
485 
486 	mutex_lock(&vnode->vfs_inode.i_mutex);
487 
488 	/* check local lock records first */
489 	ret = 0;
490 	posix_test_lock(file, fl);
491 	if (fl->fl_type == F_UNLCK) {
492 		/* no local locks; consult the server */
493 		ret = afs_vnode_fetch_status(vnode, NULL, key);
494 		if (ret < 0)
495 			goto error;
496 		lock_count = vnode->status.lock_count;
497 		if (lock_count) {
498 			if (lock_count > 0)
499 				fl->fl_type = F_RDLCK;
500 			else
501 				fl->fl_type = F_WRLCK;
502 			fl->fl_start = 0;
503 			fl->fl_end = OFFSET_MAX;
504 		}
505 	}
506 
507 error:
508 	mutex_unlock(&vnode->vfs_inode.i_mutex);
509 	_leave(" = %d [%hd]", ret, fl->fl_type);
510 	return ret;
511 }
512 
513 /*
514  * manage POSIX locks on a file
515  */
516 int afs_lock(struct file *file, int cmd, struct file_lock *fl)
517 {
518 	struct afs_vnode *vnode = AFS_FS_I(file->f_dentry->d_inode);
519 
520 	_enter("{%x:%u},%d,{t=%x,fl=%x,r=%Ld:%Ld}",
521 	       vnode->fid.vid, vnode->fid.vnode, cmd,
522 	       fl->fl_type, fl->fl_flags,
523 	       (long long) fl->fl_start, (long long) fl->fl_end);
524 
525 	/* AFS doesn't support mandatory locks */
526 	if (__mandatory_lock(&vnode->vfs_inode) && fl->fl_type != F_UNLCK)
527 		return -ENOLCK;
528 
529 	if (IS_GETLK(cmd))
530 		return afs_do_getlk(file, fl);
531 	if (fl->fl_type == F_UNLCK)
532 		return afs_do_unlk(file, fl);
533 	return afs_do_setlk(file, fl);
534 }
535 
536 /*
537  * manage FLOCK locks on a file
538  */
539 int afs_flock(struct file *file, int cmd, struct file_lock *fl)
540 {
541 	struct afs_vnode *vnode = AFS_FS_I(file->f_dentry->d_inode);
542 
543 	_enter("{%x:%u},%d,{t=%x,fl=%x}",
544 	       vnode->fid.vid, vnode->fid.vnode, cmd,
545 	       fl->fl_type, fl->fl_flags);
546 
547 	/*
548 	 * No BSD flocks over NFS allowed.
549 	 * Note: we could try to fake a POSIX lock request here by
550 	 * using ((u32) filp | 0x80000000) or some such as the pid.
551 	 * Not sure whether that would be unique, though, or whether
552 	 * that would break in other places.
553 	 */
554 	if (!(fl->fl_flags & FL_FLOCK))
555 		return -ENOLCK;
556 
557 	/* we're simulating flock() locks using posix locks on the server */
558 	fl->fl_owner = (fl_owner_t) file;
559 	fl->fl_start = 0;
560 	fl->fl_end = OFFSET_MAX;
561 
562 	if (fl->fl_type == F_UNLCK)
563 		return afs_do_unlk(file, fl);
564 	return afs_do_setlk(file, fl);
565 }
566 
567 /*
568  * the POSIX lock management core VFS code copies the lock record and adds the
569  * copy into its own list, so we need to add that copy to the vnode's lock
570  * queue in the same place as the original (which will be deleted shortly
571  * after)
572  */
573 static void afs_fl_copy_lock(struct file_lock *new, struct file_lock *fl)
574 {
575 	_enter("");
576 
577 	list_add(&new->fl_u.afs.link, &fl->fl_u.afs.link);
578 }
579 
580 /*
581  * need to remove this lock from the vnode queue when it's removed from the
582  * VFS's list
583  */
584 static void afs_fl_release_private(struct file_lock *fl)
585 {
586 	_enter("");
587 
588 	list_del_init(&fl->fl_u.afs.link);
589 }
590