Loading...
1/*
2 * linux/fs/lockd/svclock.c
3 *
4 * Handling of server-side locks, mostly of the blocked variety.
5 * This is the ugliest part of lockd because we tread on very thin ice.
6 * GRANT and CANCEL calls may get stuck, meet in mid-flight, etc.
7 * IMNSHO introducing the grant callback into the NLM protocol was one
8 * of the worst ideas Sun ever had. Except maybe for the idea of doing
9 * NFS file locking at all.
10 *
11 * I'm trying hard to avoid race conditions by protecting most accesses
12 * to a file's list of blocked locks through a semaphore. The global
13 * list of blocked locks is not protected in this fashion however.
14 * Therefore, some functions (such as the RPC callback for the async grant
15 * call) move blocked locks towards the head of the list *while some other
16 * process might be traversing it*. This should not be a problem in
17 * practice, because this will only cause functions traversing the list
18 * to visit some blocks twice.
19 *
20 * Copyright (C) 1996, Olaf Kirch <okir@monad.swb.de>
21 */
22
23#include <linux/types.h>
24#include <linux/slab.h>
25#include <linux/errno.h>
26#include <linux/kernel.h>
27#include <linux/sched.h>
28#include <linux/sunrpc/clnt.h>
29#include <linux/sunrpc/svc.h>
30#include <linux/lockd/nlm.h>
31#include <linux/lockd/lockd.h>
32#include <linux/kthread.h>
33
34#define NLMDBG_FACILITY NLMDBG_SVCLOCK
35
36#ifdef CONFIG_LOCKD_V4
37#define nlm_deadlock nlm4_deadlock
38#else
39#define nlm_deadlock nlm_lck_denied
40#endif
41
42static void nlmsvc_release_block(struct nlm_block *block);
43static void nlmsvc_insert_block(struct nlm_block *block, unsigned long);
44static void nlmsvc_remove_block(struct nlm_block *block);
45
46static int nlmsvc_setgrantargs(struct nlm_rqst *call, struct nlm_lock *lock);
47static void nlmsvc_freegrantargs(struct nlm_rqst *call);
48static const struct rpc_call_ops nlmsvc_grant_ops;
49static const char *nlmdbg_cookie2a(const struct nlm_cookie *cookie);
50
51/*
52 * The list of blocked locks to retry
53 */
54static LIST_HEAD(nlm_blocked);
55static DEFINE_SPINLOCK(nlm_blocked_lock);
56
57/*
58 * Insert a blocked lock into the global list
59 */
60static void
61nlmsvc_insert_block_locked(struct nlm_block *block, unsigned long when)
62{
63 struct nlm_block *b;
64 struct list_head *pos;
65
66 dprintk("lockd: nlmsvc_insert_block(%p, %ld)\n", block, when);
67 if (list_empty(&block->b_list)) {
68 kref_get(&block->b_count);
69 } else {
70 list_del_init(&block->b_list);
71 }
72
73 pos = &nlm_blocked;
74 if (when != NLM_NEVER) {
75 if ((when += jiffies) == NLM_NEVER)
76 when ++;
77 list_for_each(pos, &nlm_blocked) {
78 b = list_entry(pos, struct nlm_block, b_list);
79 if (time_after(b->b_when,when) || b->b_when == NLM_NEVER)
80 break;
81 }
82 /* On normal exit from the loop, pos == &nlm_blocked,
83 * so we will be adding to the end of the list - good
84 */
85 }
86
87 list_add_tail(&block->b_list, pos);
88 block->b_when = when;
89}
90
91static void nlmsvc_insert_block(struct nlm_block *block, unsigned long when)
92{
93 spin_lock(&nlm_blocked_lock);
94 nlmsvc_insert_block_locked(block, when);
95 spin_unlock(&nlm_blocked_lock);
96}
97
98/*
99 * Remove a block from the global list
100 */
101static inline void
102nlmsvc_remove_block(struct nlm_block *block)
103{
104 if (!list_empty(&block->b_list)) {
105 spin_lock(&nlm_blocked_lock);
106 list_del_init(&block->b_list);
107 spin_unlock(&nlm_blocked_lock);
108 nlmsvc_release_block(block);
109 }
110}
111
112/*
113 * Find a block for a given lock
114 */
115static struct nlm_block *
116nlmsvc_lookup_block(struct nlm_file *file, struct nlm_lock *lock)
117{
118 struct nlm_block *block;
119 struct file_lock *fl;
120
121 dprintk("lockd: nlmsvc_lookup_block f=%p pd=%d %Ld-%Ld ty=%d\n",
122 file, lock->fl.fl_pid,
123 (long long)lock->fl.fl_start,
124 (long long)lock->fl.fl_end, lock->fl.fl_type);
125 list_for_each_entry(block, &nlm_blocked, b_list) {
126 fl = &block->b_call->a_args.lock.fl;
127 dprintk("lockd: check f=%p pd=%d %Ld-%Ld ty=%d cookie=%s\n",
128 block->b_file, fl->fl_pid,
129 (long long)fl->fl_start,
130 (long long)fl->fl_end, fl->fl_type,
131 nlmdbg_cookie2a(&block->b_call->a_args.cookie));
132 if (block->b_file == file && nlm_compare_locks(fl, &lock->fl)) {
133 kref_get(&block->b_count);
134 return block;
135 }
136 }
137
138 return NULL;
139}
140
141static inline int nlm_cookie_match(struct nlm_cookie *a, struct nlm_cookie *b)
142{
143 if (a->len != b->len)
144 return 0;
145 if (memcmp(a->data, b->data, a->len))
146 return 0;
147 return 1;
148}
149
150/*
151 * Find a block with a given NLM cookie.
152 */
153static inline struct nlm_block *
154nlmsvc_find_block(struct nlm_cookie *cookie)
155{
156 struct nlm_block *block;
157
158 list_for_each_entry(block, &nlm_blocked, b_list) {
159 if (nlm_cookie_match(&block->b_call->a_args.cookie,cookie))
160 goto found;
161 }
162
163 return NULL;
164
165found:
166 dprintk("nlmsvc_find_block(%s): block=%p\n", nlmdbg_cookie2a(cookie), block);
167 kref_get(&block->b_count);
168 return block;
169}
170
171/*
172 * Create a block and initialize it.
173 *
174 * Note: we explicitly set the cookie of the grant reply to that of
175 * the blocked lock request. The spec explicitly mentions that the client
176 * should _not_ rely on the callback containing the same cookie as the
177 * request, but (as I found out later) that's because some implementations
178 * do just this. Never mind the standards comittees, they support our
179 * logging industries.
180 *
181 * 10 years later: I hope we can safely ignore these old and broken
182 * clients by now. Let's fix this so we can uniquely identify an incoming
183 * GRANTED_RES message by cookie, without having to rely on the client's IP
184 * address. --okir
185 */
186static struct nlm_block *
187nlmsvc_create_block(struct svc_rqst *rqstp, struct nlm_host *host,
188 struct nlm_file *file, struct nlm_lock *lock,
189 struct nlm_cookie *cookie)
190{
191 struct nlm_block *block;
192 struct nlm_rqst *call = NULL;
193
194 nlm_get_host(host);
195 call = nlm_alloc_call(host);
196 if (call == NULL)
197 return NULL;
198
199 /* Allocate memory for block, and initialize arguments */
200 block = kzalloc(sizeof(*block), GFP_KERNEL);
201 if (block == NULL)
202 goto failed;
203 kref_init(&block->b_count);
204 INIT_LIST_HEAD(&block->b_list);
205 INIT_LIST_HEAD(&block->b_flist);
206
207 if (!nlmsvc_setgrantargs(call, lock))
208 goto failed_free;
209
210 /* Set notifier function for VFS, and init args */
211 call->a_args.lock.fl.fl_flags |= FL_SLEEP;
212 call->a_args.lock.fl.fl_lmops = &nlmsvc_lock_operations;
213 nlmclnt_next_cookie(&call->a_args.cookie);
214
215 dprintk("lockd: created block %p...\n", block);
216
217 /* Create and initialize the block */
218 block->b_daemon = rqstp->rq_server;
219 block->b_host = host;
220 block->b_file = file;
221 block->b_fl = NULL;
222 file->f_count++;
223
224 /* Add to file's list of blocks */
225 list_add(&block->b_flist, &file->f_blocks);
226
227 /* Set up RPC arguments for callback */
228 block->b_call = call;
229 call->a_flags = RPC_TASK_ASYNC;
230 call->a_block = block;
231
232 return block;
233
234failed_free:
235 kfree(block);
236failed:
237 nlmsvc_release_call(call);
238 return NULL;
239}
240
241/*
242 * Delete a block.
243 * It is the caller's responsibility to check whether the file
244 * can be closed hereafter.
245 */
246static int nlmsvc_unlink_block(struct nlm_block *block)
247{
248 int status;
249 dprintk("lockd: unlinking block %p...\n", block);
250
251 /* Remove block from list */
252 status = posix_unblock_lock(block->b_file->f_file, &block->b_call->a_args.lock.fl);
253 nlmsvc_remove_block(block);
254 return status;
255}
256
257static void nlmsvc_free_block(struct kref *kref)
258{
259 struct nlm_block *block = container_of(kref, struct nlm_block, b_count);
260 struct nlm_file *file = block->b_file;
261
262 dprintk("lockd: freeing block %p...\n", block);
263
264 /* Remove block from file's list of blocks */
265 mutex_lock(&file->f_mutex);
266 list_del_init(&block->b_flist);
267 mutex_unlock(&file->f_mutex);
268
269 nlmsvc_freegrantargs(block->b_call);
270 nlmsvc_release_call(block->b_call);
271 nlm_release_file(block->b_file);
272 kfree(block->b_fl);
273 kfree(block);
274}
275
276static void nlmsvc_release_block(struct nlm_block *block)
277{
278 if (block != NULL)
279 kref_put(&block->b_count, nlmsvc_free_block);
280}
281
282/*
283 * Loop over all blocks and delete blocks held by
284 * a matching host.
285 */
286void nlmsvc_traverse_blocks(struct nlm_host *host,
287 struct nlm_file *file,
288 nlm_host_match_fn_t match)
289{
290 struct nlm_block *block, *next;
291
292restart:
293 mutex_lock(&file->f_mutex);
294 list_for_each_entry_safe(block, next, &file->f_blocks, b_flist) {
295 if (!match(block->b_host, host))
296 continue;
297 /* Do not destroy blocks that are not on
298 * the global retry list - why? */
299 if (list_empty(&block->b_list))
300 continue;
301 kref_get(&block->b_count);
302 mutex_unlock(&file->f_mutex);
303 nlmsvc_unlink_block(block);
304 nlmsvc_release_block(block);
305 goto restart;
306 }
307 mutex_unlock(&file->f_mutex);
308}
309
310/*
311 * Initialize arguments for GRANTED call. The nlm_rqst structure
312 * has been cleared already.
313 */
314static int nlmsvc_setgrantargs(struct nlm_rqst *call, struct nlm_lock *lock)
315{
316 locks_copy_lock(&call->a_args.lock.fl, &lock->fl);
317 memcpy(&call->a_args.lock.fh, &lock->fh, sizeof(call->a_args.lock.fh));
318 call->a_args.lock.caller = utsname()->nodename;
319 call->a_args.lock.oh.len = lock->oh.len;
320
321 /* set default data area */
322 call->a_args.lock.oh.data = call->a_owner;
323 call->a_args.lock.svid = lock->fl.fl_pid;
324
325 if (lock->oh.len > NLMCLNT_OHSIZE) {
326 void *data = kmalloc(lock->oh.len, GFP_KERNEL);
327 if (!data)
328 return 0;
329 call->a_args.lock.oh.data = (u8 *) data;
330 }
331
332 memcpy(call->a_args.lock.oh.data, lock->oh.data, lock->oh.len);
333 return 1;
334}
335
336static void nlmsvc_freegrantargs(struct nlm_rqst *call)
337{
338 if (call->a_args.lock.oh.data != call->a_owner)
339 kfree(call->a_args.lock.oh.data);
340
341 locks_release_private(&call->a_args.lock.fl);
342}
343
344/*
345 * Deferred lock request handling for non-blocking lock
346 */
347static __be32
348nlmsvc_defer_lock_rqst(struct svc_rqst *rqstp, struct nlm_block *block)
349{
350 __be32 status = nlm_lck_denied_nolocks;
351
352 block->b_flags |= B_QUEUED;
353
354 nlmsvc_insert_block(block, NLM_TIMEOUT);
355
356 block->b_cache_req = &rqstp->rq_chandle;
357 if (rqstp->rq_chandle.defer) {
358 block->b_deferred_req =
359 rqstp->rq_chandle.defer(block->b_cache_req);
360 if (block->b_deferred_req != NULL)
361 status = nlm_drop_reply;
362 }
363 dprintk("lockd: nlmsvc_defer_lock_rqst block %p flags %d status %d\n",
364 block, block->b_flags, ntohl(status));
365
366 return status;
367}
368
369/*
370 * Attempt to establish a lock, and if it can't be granted, block it
371 * if required.
372 */
373__be32
374nlmsvc_lock(struct svc_rqst *rqstp, struct nlm_file *file,
375 struct nlm_host *host, struct nlm_lock *lock, int wait,
376 struct nlm_cookie *cookie, int reclaim)
377{
378 struct nlm_block *block = NULL;
379 int error;
380 __be32 ret;
381
382 dprintk("lockd: nlmsvc_lock(%s/%ld, ty=%d, pi=%d, %Ld-%Ld, bl=%d)\n",
383 file->f_file->f_path.dentry->d_inode->i_sb->s_id,
384 file->f_file->f_path.dentry->d_inode->i_ino,
385 lock->fl.fl_type, lock->fl.fl_pid,
386 (long long)lock->fl.fl_start,
387 (long long)lock->fl.fl_end,
388 wait);
389
390 /* Lock file against concurrent access */
391 mutex_lock(&file->f_mutex);
392 /* Get existing block (in case client is busy-waiting)
393 * or create new block
394 */
395 block = nlmsvc_lookup_block(file, lock);
396 if (block == NULL) {
397 block = nlmsvc_create_block(rqstp, host, file, lock, cookie);
398 ret = nlm_lck_denied_nolocks;
399 if (block == NULL)
400 goto out;
401 lock = &block->b_call->a_args.lock;
402 } else
403 lock->fl.fl_flags &= ~FL_SLEEP;
404
405 if (block->b_flags & B_QUEUED) {
406 dprintk("lockd: nlmsvc_lock deferred block %p flags %d\n",
407 block, block->b_flags);
408 if (block->b_granted) {
409 nlmsvc_unlink_block(block);
410 ret = nlm_granted;
411 goto out;
412 }
413 if (block->b_flags & B_TIMED_OUT) {
414 nlmsvc_unlink_block(block);
415 ret = nlm_lck_denied;
416 goto out;
417 }
418 ret = nlm_drop_reply;
419 goto out;
420 }
421
422 if (locks_in_grace() && !reclaim) {
423 ret = nlm_lck_denied_grace_period;
424 goto out;
425 }
426 if (reclaim && !locks_in_grace()) {
427 ret = nlm_lck_denied_grace_period;
428 goto out;
429 }
430
431 if (!wait)
432 lock->fl.fl_flags &= ~FL_SLEEP;
433 error = vfs_lock_file(file->f_file, F_SETLK, &lock->fl, NULL);
434 lock->fl.fl_flags &= ~FL_SLEEP;
435
436 dprintk("lockd: vfs_lock_file returned %d\n", error);
437 switch (error) {
438 case 0:
439 ret = nlm_granted;
440 goto out;
441 case -EAGAIN:
442 /*
443 * If this is a blocking request for an
444 * already pending lock request then we need
445 * to put it back on lockd's block list
446 */
447 if (wait)
448 break;
449 ret = nlm_lck_denied;
450 goto out;
451 case FILE_LOCK_DEFERRED:
452 if (wait)
453 break;
454 /* Filesystem lock operation is in progress
455 Add it to the queue waiting for callback */
456 ret = nlmsvc_defer_lock_rqst(rqstp, block);
457 goto out;
458 case -EDEADLK:
459 ret = nlm_deadlock;
460 goto out;
461 default: /* includes ENOLCK */
462 ret = nlm_lck_denied_nolocks;
463 goto out;
464 }
465
466 ret = nlm_lck_blocked;
467
468 /* Append to list of blocked */
469 nlmsvc_insert_block(block, NLM_NEVER);
470out:
471 mutex_unlock(&file->f_mutex);
472 nlmsvc_release_block(block);
473 dprintk("lockd: nlmsvc_lock returned %u\n", ret);
474 return ret;
475}
476
477/*
478 * Test for presence of a conflicting lock.
479 */
480__be32
481nlmsvc_testlock(struct svc_rqst *rqstp, struct nlm_file *file,
482 struct nlm_host *host, struct nlm_lock *lock,
483 struct nlm_lock *conflock, struct nlm_cookie *cookie)
484{
485 struct nlm_block *block = NULL;
486 int error;
487 __be32 ret;
488
489 dprintk("lockd: nlmsvc_testlock(%s/%ld, ty=%d, %Ld-%Ld)\n",
490 file->f_file->f_path.dentry->d_inode->i_sb->s_id,
491 file->f_file->f_path.dentry->d_inode->i_ino,
492 lock->fl.fl_type,
493 (long long)lock->fl.fl_start,
494 (long long)lock->fl.fl_end);
495
496 /* Get existing block (in case client is busy-waiting) */
497 block = nlmsvc_lookup_block(file, lock);
498
499 if (block == NULL) {
500 struct file_lock *conf = kzalloc(sizeof(*conf), GFP_KERNEL);
501
502 if (conf == NULL)
503 return nlm_granted;
504 block = nlmsvc_create_block(rqstp, host, file, lock, cookie);
505 if (block == NULL) {
506 kfree(conf);
507 return nlm_granted;
508 }
509 block->b_fl = conf;
510 }
511 if (block->b_flags & B_QUEUED) {
512 dprintk("lockd: nlmsvc_testlock deferred block %p flags %d fl %p\n",
513 block, block->b_flags, block->b_fl);
514 if (block->b_flags & B_TIMED_OUT) {
515 nlmsvc_unlink_block(block);
516 ret = nlm_lck_denied;
517 goto out;
518 }
519 if (block->b_flags & B_GOT_CALLBACK) {
520 nlmsvc_unlink_block(block);
521 if (block->b_fl != NULL
522 && block->b_fl->fl_type != F_UNLCK) {
523 lock->fl = *block->b_fl;
524 goto conf_lock;
525 } else {
526 ret = nlm_granted;
527 goto out;
528 }
529 }
530 ret = nlm_drop_reply;
531 goto out;
532 }
533
534 if (locks_in_grace()) {
535 ret = nlm_lck_denied_grace_period;
536 goto out;
537 }
538 error = vfs_test_lock(file->f_file, &lock->fl);
539 if (error == FILE_LOCK_DEFERRED) {
540 ret = nlmsvc_defer_lock_rqst(rqstp, block);
541 goto out;
542 }
543 if (error) {
544 ret = nlm_lck_denied_nolocks;
545 goto out;
546 }
547 if (lock->fl.fl_type == F_UNLCK) {
548 ret = nlm_granted;
549 goto out;
550 }
551
552conf_lock:
553 dprintk("lockd: conflicting lock(ty=%d, %Ld-%Ld)\n",
554 lock->fl.fl_type, (long long)lock->fl.fl_start,
555 (long long)lock->fl.fl_end);
556 conflock->caller = "somehost"; /* FIXME */
557 conflock->len = strlen(conflock->caller);
558 conflock->oh.len = 0; /* don't return OH info */
559 conflock->svid = lock->fl.fl_pid;
560 conflock->fl.fl_type = lock->fl.fl_type;
561 conflock->fl.fl_start = lock->fl.fl_start;
562 conflock->fl.fl_end = lock->fl.fl_end;
563 ret = nlm_lck_denied;
564out:
565 if (block)
566 nlmsvc_release_block(block);
567 return ret;
568}
569
570/*
571 * Remove a lock.
572 * This implies a CANCEL call: We send a GRANT_MSG, the client replies
573 * with a GRANT_RES call which gets lost, and calls UNLOCK immediately
574 * afterwards. In this case the block will still be there, and hence
575 * must be removed.
576 */
577__be32
578nlmsvc_unlock(struct nlm_file *file, struct nlm_lock *lock)
579{
580 int error;
581
582 dprintk("lockd: nlmsvc_unlock(%s/%ld, pi=%d, %Ld-%Ld)\n",
583 file->f_file->f_path.dentry->d_inode->i_sb->s_id,
584 file->f_file->f_path.dentry->d_inode->i_ino,
585 lock->fl.fl_pid,
586 (long long)lock->fl.fl_start,
587 (long long)lock->fl.fl_end);
588
589 /* First, cancel any lock that might be there */
590 nlmsvc_cancel_blocked(file, lock);
591
592 lock->fl.fl_type = F_UNLCK;
593 error = vfs_lock_file(file->f_file, F_SETLK, &lock->fl, NULL);
594
595 return (error < 0)? nlm_lck_denied_nolocks : nlm_granted;
596}
597
598/*
599 * Cancel a previously blocked request.
600 *
601 * A cancel request always overrides any grant that may currently
602 * be in progress.
603 * The calling procedure must check whether the file can be closed.
604 */
605__be32
606nlmsvc_cancel_blocked(struct nlm_file *file, struct nlm_lock *lock)
607{
608 struct nlm_block *block;
609 int status = 0;
610
611 dprintk("lockd: nlmsvc_cancel(%s/%ld, pi=%d, %Ld-%Ld)\n",
612 file->f_file->f_path.dentry->d_inode->i_sb->s_id,
613 file->f_file->f_path.dentry->d_inode->i_ino,
614 lock->fl.fl_pid,
615 (long long)lock->fl.fl_start,
616 (long long)lock->fl.fl_end);
617
618 if (locks_in_grace())
619 return nlm_lck_denied_grace_period;
620
621 mutex_lock(&file->f_mutex);
622 block = nlmsvc_lookup_block(file, lock);
623 mutex_unlock(&file->f_mutex);
624 if (block != NULL) {
625 vfs_cancel_lock(block->b_file->f_file,
626 &block->b_call->a_args.lock.fl);
627 status = nlmsvc_unlink_block(block);
628 nlmsvc_release_block(block);
629 }
630 return status ? nlm_lck_denied : nlm_granted;
631}
632
633/*
634 * This is a callback from the filesystem for VFS file lock requests.
635 * It will be used if lm_grant is defined and the filesystem can not
636 * respond to the request immediately.
637 * For GETLK request it will copy the reply to the nlm_block.
638 * For SETLK or SETLKW request it will get the local posix lock.
639 * In all cases it will move the block to the head of nlm_blocked q where
640 * nlmsvc_retry_blocked() can send back a reply for SETLKW or revisit the
641 * deferred rpc for GETLK and SETLK.
642 */
643static void
644nlmsvc_update_deferred_block(struct nlm_block *block, struct file_lock *conf,
645 int result)
646{
647 block->b_flags |= B_GOT_CALLBACK;
648 if (result == 0)
649 block->b_granted = 1;
650 else
651 block->b_flags |= B_TIMED_OUT;
652 if (conf) {
653 if (block->b_fl)
654 __locks_copy_lock(block->b_fl, conf);
655 }
656}
657
658static int nlmsvc_grant_deferred(struct file_lock *fl, struct file_lock *conf,
659 int result)
660{
661 struct nlm_block *block;
662 int rc = -ENOENT;
663
664 spin_lock(&nlm_blocked_lock);
665 list_for_each_entry(block, &nlm_blocked, b_list) {
666 if (nlm_compare_locks(&block->b_call->a_args.lock.fl, fl)) {
667 dprintk("lockd: nlmsvc_notify_blocked block %p flags %d\n",
668 block, block->b_flags);
669 if (block->b_flags & B_QUEUED) {
670 if (block->b_flags & B_TIMED_OUT) {
671 rc = -ENOLCK;
672 break;
673 }
674 nlmsvc_update_deferred_block(block, conf, result);
675 } else if (result == 0)
676 block->b_granted = 1;
677
678 nlmsvc_insert_block_locked(block, 0);
679 svc_wake_up(block->b_daemon);
680 rc = 0;
681 break;
682 }
683 }
684 spin_unlock(&nlm_blocked_lock);
685 if (rc == -ENOENT)
686 printk(KERN_WARNING "lockd: grant for unknown block\n");
687 return rc;
688}
689
690/*
691 * Unblock a blocked lock request. This is a callback invoked from the
692 * VFS layer when a lock on which we blocked is removed.
693 *
694 * This function doesn't grant the blocked lock instantly, but rather moves
695 * the block to the head of nlm_blocked where it can be picked up by lockd.
696 */
697static void
698nlmsvc_notify_blocked(struct file_lock *fl)
699{
700 struct nlm_block *block;
701
702 dprintk("lockd: VFS unblock notification for block %p\n", fl);
703 spin_lock(&nlm_blocked_lock);
704 list_for_each_entry(block, &nlm_blocked, b_list) {
705 if (nlm_compare_locks(&block->b_call->a_args.lock.fl, fl)) {
706 nlmsvc_insert_block_locked(block, 0);
707 spin_unlock(&nlm_blocked_lock);
708 svc_wake_up(block->b_daemon);
709 return;
710 }
711 }
712 spin_unlock(&nlm_blocked_lock);
713 printk(KERN_WARNING "lockd: notification for unknown block!\n");
714}
715
716static int nlmsvc_same_owner(struct file_lock *fl1, struct file_lock *fl2)
717{
718 return fl1->fl_owner == fl2->fl_owner && fl1->fl_pid == fl2->fl_pid;
719}
720
721const struct lock_manager_operations nlmsvc_lock_operations = {
722 .lm_compare_owner = nlmsvc_same_owner,
723 .lm_notify = nlmsvc_notify_blocked,
724 .lm_grant = nlmsvc_grant_deferred,
725};
726
727/*
728 * Try to claim a lock that was previously blocked.
729 *
730 * Note that we use both the RPC_GRANTED_MSG call _and_ an async
731 * RPC thread when notifying the client. This seems like overkill...
732 * Here's why:
733 * - we don't want to use a synchronous RPC thread, otherwise
734 * we might find ourselves hanging on a dead portmapper.
735 * - Some lockd implementations (e.g. HP) don't react to
736 * RPC_GRANTED calls; they seem to insist on RPC_GRANTED_MSG calls.
737 */
738static void
739nlmsvc_grant_blocked(struct nlm_block *block)
740{
741 struct nlm_file *file = block->b_file;
742 struct nlm_lock *lock = &block->b_call->a_args.lock;
743 int error;
744
745 dprintk("lockd: grant blocked lock %p\n", block);
746
747 kref_get(&block->b_count);
748
749 /* Unlink block request from list */
750 nlmsvc_unlink_block(block);
751
752 /* If b_granted is true this means we've been here before.
753 * Just retry the grant callback, possibly refreshing the RPC
754 * binding */
755 if (block->b_granted) {
756 nlm_rebind_host(block->b_host);
757 goto callback;
758 }
759
760 /* Try the lock operation again */
761 lock->fl.fl_flags |= FL_SLEEP;
762 error = vfs_lock_file(file->f_file, F_SETLK, &lock->fl, NULL);
763 lock->fl.fl_flags &= ~FL_SLEEP;
764
765 switch (error) {
766 case 0:
767 break;
768 case FILE_LOCK_DEFERRED:
769 dprintk("lockd: lock still blocked error %d\n", error);
770 nlmsvc_insert_block(block, NLM_NEVER);
771 nlmsvc_release_block(block);
772 return;
773 default:
774 printk(KERN_WARNING "lockd: unexpected error %d in %s!\n",
775 -error, __func__);
776 nlmsvc_insert_block(block, 10 * HZ);
777 nlmsvc_release_block(block);
778 return;
779 }
780
781callback:
782 /* Lock was granted by VFS. */
783 dprintk("lockd: GRANTing blocked lock.\n");
784 block->b_granted = 1;
785
786 /* keep block on the list, but don't reattempt until the RPC
787 * completes or the submission fails
788 */
789 nlmsvc_insert_block(block, NLM_NEVER);
790
791 /* Call the client -- use a soft RPC task since nlmsvc_retry_blocked
792 * will queue up a new one if this one times out
793 */
794 error = nlm_async_call(block->b_call, NLMPROC_GRANTED_MSG,
795 &nlmsvc_grant_ops);
796
797 /* RPC submission failed, wait a bit and retry */
798 if (error < 0)
799 nlmsvc_insert_block(block, 10 * HZ);
800}
801
802/*
803 * This is the callback from the RPC layer when the NLM_GRANTED_MSG
804 * RPC call has succeeded or timed out.
805 * Like all RPC callbacks, it is invoked by the rpciod process, so it
806 * better not sleep. Therefore, we put the blocked lock on the nlm_blocked
807 * chain once more in order to have it removed by lockd itself (which can
808 * then sleep on the file semaphore without disrupting e.g. the nfs client).
809 */
810static void nlmsvc_grant_callback(struct rpc_task *task, void *data)
811{
812 struct nlm_rqst *call = data;
813 struct nlm_block *block = call->a_block;
814 unsigned long timeout;
815
816 dprintk("lockd: GRANT_MSG RPC callback\n");
817
818 spin_lock(&nlm_blocked_lock);
819 /* if the block is not on a list at this point then it has
820 * been invalidated. Don't try to requeue it.
821 *
822 * FIXME: it's possible that the block is removed from the list
823 * after this check but before the nlmsvc_insert_block. In that
824 * case it will be added back. Perhaps we need better locking
825 * for nlm_blocked?
826 */
827 if (list_empty(&block->b_list))
828 goto out;
829
830 /* Technically, we should down the file semaphore here. Since we
831 * move the block towards the head of the queue only, no harm
832 * can be done, though. */
833 if (task->tk_status < 0) {
834 /* RPC error: Re-insert for retransmission */
835 timeout = 10 * HZ;
836 } else {
837 /* Call was successful, now wait for client callback */
838 timeout = 60 * HZ;
839 }
840 nlmsvc_insert_block_locked(block, timeout);
841 svc_wake_up(block->b_daemon);
842out:
843 spin_unlock(&nlm_blocked_lock);
844}
845
846/*
847 * FIXME: nlmsvc_release_block() grabs a mutex. This is not allowed for an
848 * .rpc_release rpc_call_op
849 */
850static void nlmsvc_grant_release(void *data)
851{
852 struct nlm_rqst *call = data;
853 nlmsvc_release_block(call->a_block);
854}
855
856static const struct rpc_call_ops nlmsvc_grant_ops = {
857 .rpc_call_done = nlmsvc_grant_callback,
858 .rpc_release = nlmsvc_grant_release,
859};
860
861/*
862 * We received a GRANT_RES callback. Try to find the corresponding
863 * block.
864 */
865void
866nlmsvc_grant_reply(struct nlm_cookie *cookie, __be32 status)
867{
868 struct nlm_block *block;
869
870 dprintk("grant_reply: looking for cookie %x, s=%d \n",
871 *(unsigned int *)(cookie->data), status);
872 if (!(block = nlmsvc_find_block(cookie)))
873 return;
874
875 if (block) {
876 if (status == nlm_lck_denied_grace_period) {
877 /* Try again in a couple of seconds */
878 nlmsvc_insert_block(block, 10 * HZ);
879 } else {
880 /* Lock is now held by client, or has been rejected.
881 * In both cases, the block should be removed. */
882 nlmsvc_unlink_block(block);
883 }
884 }
885 nlmsvc_release_block(block);
886}
887
888/* Helper function to handle retry of a deferred block.
889 * If it is a blocking lock, call grant_blocked.
890 * For a non-blocking lock or test lock, revisit the request.
891 */
892static void
893retry_deferred_block(struct nlm_block *block)
894{
895 if (!(block->b_flags & B_GOT_CALLBACK))
896 block->b_flags |= B_TIMED_OUT;
897 nlmsvc_insert_block(block, NLM_TIMEOUT);
898 dprintk("revisit block %p flags %d\n", block, block->b_flags);
899 if (block->b_deferred_req) {
900 block->b_deferred_req->revisit(block->b_deferred_req, 0);
901 block->b_deferred_req = NULL;
902 }
903}
904
905/*
906 * Retry all blocked locks that have been notified. This is where lockd
907 * picks up locks that can be granted, or grant notifications that must
908 * be retransmitted.
909 */
910unsigned long
911nlmsvc_retry_blocked(void)
912{
913 unsigned long timeout = MAX_SCHEDULE_TIMEOUT;
914 struct nlm_block *block;
915
916 while (!list_empty(&nlm_blocked) && !kthread_should_stop()) {
917 block = list_entry(nlm_blocked.next, struct nlm_block, b_list);
918
919 if (block->b_when == NLM_NEVER)
920 break;
921 if (time_after(block->b_when, jiffies)) {
922 timeout = block->b_when - jiffies;
923 break;
924 }
925
926 dprintk("nlmsvc_retry_blocked(%p, when=%ld)\n",
927 block, block->b_when);
928 if (block->b_flags & B_QUEUED) {
929 dprintk("nlmsvc_retry_blocked delete block (%p, granted=%d, flags=%d)\n",
930 block, block->b_granted, block->b_flags);
931 retry_deferred_block(block);
932 } else
933 nlmsvc_grant_blocked(block);
934 }
935
936 return timeout;
937}
938
939#ifdef RPC_DEBUG
940static const char *nlmdbg_cookie2a(const struct nlm_cookie *cookie)
941{
942 /*
943 * We can get away with a static buffer because we're only
944 * called with BKL held.
945 */
946 static char buf[2*NLM_MAXCOOKIELEN+1];
947 unsigned int i, len = sizeof(buf);
948 char *p = buf;
949
950 len--; /* allow for trailing \0 */
951 if (len < 3)
952 return "???";
953 for (i = 0 ; i < cookie->len ; i++) {
954 if (len < 2) {
955 strcpy(p-3, "...");
956 break;
957 }
958 sprintf(p, "%02x", cookie->data[i]);
959 p += 2;
960 len -= 2;
961 }
962 *p = '\0';
963
964 return buf;
965}
966#endif
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * linux/fs/lockd/svclock.c
4 *
5 * Handling of server-side locks, mostly of the blocked variety.
6 * This is the ugliest part of lockd because we tread on very thin ice.
7 * GRANT and CANCEL calls may get stuck, meet in mid-flight, etc.
8 * IMNSHO introducing the grant callback into the NLM protocol was one
9 * of the worst ideas Sun ever had. Except maybe for the idea of doing
10 * NFS file locking at all.
11 *
12 * I'm trying hard to avoid race conditions by protecting most accesses
13 * to a file's list of blocked locks through a semaphore. The global
14 * list of blocked locks is not protected in this fashion however.
15 * Therefore, some functions (such as the RPC callback for the async grant
16 * call) move blocked locks towards the head of the list *while some other
17 * process might be traversing it*. This should not be a problem in
18 * practice, because this will only cause functions traversing the list
19 * to visit some blocks twice.
20 *
21 * Copyright (C) 1996, Olaf Kirch <okir@monad.swb.de>
22 */
23
24#include <linux/types.h>
25#include <linux/slab.h>
26#include <linux/errno.h>
27#include <linux/kernel.h>
28#include <linux/sched.h>
29#include <linux/sunrpc/clnt.h>
30#include <linux/sunrpc/svc_xprt.h>
31#include <linux/lockd/nlm.h>
32#include <linux/lockd/lockd.h>
33#include <linux/kthread.h>
34
35#define NLMDBG_FACILITY NLMDBG_SVCLOCK
36
37#ifdef CONFIG_LOCKD_V4
38#define nlm_deadlock nlm4_deadlock
39#else
40#define nlm_deadlock nlm_lck_denied
41#endif
42
43static void nlmsvc_release_block(struct nlm_block *block);
44static void nlmsvc_insert_block(struct nlm_block *block, unsigned long);
45static void nlmsvc_remove_block(struct nlm_block *block);
46
47static int nlmsvc_setgrantargs(struct nlm_rqst *call, struct nlm_lock *lock);
48static void nlmsvc_freegrantargs(struct nlm_rqst *call);
49static const struct rpc_call_ops nlmsvc_grant_ops;
50
51/*
52 * The list of blocked locks to retry
53 */
54static LIST_HEAD(nlm_blocked);
55static DEFINE_SPINLOCK(nlm_blocked_lock);
56
57#if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
58static const char *nlmdbg_cookie2a(const struct nlm_cookie *cookie)
59{
60 /*
61 * We can get away with a static buffer because this is only called
62 * from lockd, which is single-threaded.
63 */
64 static char buf[2*NLM_MAXCOOKIELEN+1];
65 unsigned int i, len = sizeof(buf);
66 char *p = buf;
67
68 len--; /* allow for trailing \0 */
69 if (len < 3)
70 return "???";
71 for (i = 0 ; i < cookie->len ; i++) {
72 if (len < 2) {
73 strcpy(p-3, "...");
74 break;
75 }
76 sprintf(p, "%02x", cookie->data[i]);
77 p += 2;
78 len -= 2;
79 }
80 *p = '\0';
81
82 return buf;
83}
84#endif
85
86/*
87 * Insert a blocked lock into the global list
88 */
89static void
90nlmsvc_insert_block_locked(struct nlm_block *block, unsigned long when)
91{
92 struct nlm_block *b;
93 struct list_head *pos;
94
95 dprintk("lockd: nlmsvc_insert_block(%p, %ld)\n", block, when);
96 if (list_empty(&block->b_list)) {
97 kref_get(&block->b_count);
98 } else {
99 list_del_init(&block->b_list);
100 }
101
102 pos = &nlm_blocked;
103 if (when != NLM_NEVER) {
104 if ((when += jiffies) == NLM_NEVER)
105 when ++;
106 list_for_each(pos, &nlm_blocked) {
107 b = list_entry(pos, struct nlm_block, b_list);
108 if (time_after(b->b_when,when) || b->b_when == NLM_NEVER)
109 break;
110 }
111 /* On normal exit from the loop, pos == &nlm_blocked,
112 * so we will be adding to the end of the list - good
113 */
114 }
115
116 list_add_tail(&block->b_list, pos);
117 block->b_when = when;
118}
119
120static void nlmsvc_insert_block(struct nlm_block *block, unsigned long when)
121{
122 spin_lock(&nlm_blocked_lock);
123 nlmsvc_insert_block_locked(block, when);
124 spin_unlock(&nlm_blocked_lock);
125}
126
127/*
128 * Remove a block from the global list
129 */
130static inline void
131nlmsvc_remove_block(struct nlm_block *block)
132{
133 if (!list_empty(&block->b_list)) {
134 spin_lock(&nlm_blocked_lock);
135 list_del_init(&block->b_list);
136 spin_unlock(&nlm_blocked_lock);
137 nlmsvc_release_block(block);
138 }
139}
140
141/*
142 * Find a block for a given lock
143 */
144static struct nlm_block *
145nlmsvc_lookup_block(struct nlm_file *file, struct nlm_lock *lock)
146{
147 struct nlm_block *block;
148 struct file_lock *fl;
149
150 dprintk("lockd: nlmsvc_lookup_block f=%p pd=%d %Ld-%Ld ty=%d\n",
151 file, lock->fl.fl_pid,
152 (long long)lock->fl.fl_start,
153 (long long)lock->fl.fl_end, lock->fl.fl_type);
154 list_for_each_entry(block, &nlm_blocked, b_list) {
155 fl = &block->b_call->a_args.lock.fl;
156 dprintk("lockd: check f=%p pd=%d %Ld-%Ld ty=%d cookie=%s\n",
157 block->b_file, fl->fl_pid,
158 (long long)fl->fl_start,
159 (long long)fl->fl_end, fl->fl_type,
160 nlmdbg_cookie2a(&block->b_call->a_args.cookie));
161 if (block->b_file == file && nlm_compare_locks(fl, &lock->fl)) {
162 kref_get(&block->b_count);
163 return block;
164 }
165 }
166
167 return NULL;
168}
169
170static inline int nlm_cookie_match(struct nlm_cookie *a, struct nlm_cookie *b)
171{
172 if (a->len != b->len)
173 return 0;
174 if (memcmp(a->data, b->data, a->len))
175 return 0;
176 return 1;
177}
178
179/*
180 * Find a block with a given NLM cookie.
181 */
182static inline struct nlm_block *
183nlmsvc_find_block(struct nlm_cookie *cookie)
184{
185 struct nlm_block *block;
186
187 list_for_each_entry(block, &nlm_blocked, b_list) {
188 if (nlm_cookie_match(&block->b_call->a_args.cookie,cookie))
189 goto found;
190 }
191
192 return NULL;
193
194found:
195 dprintk("nlmsvc_find_block(%s): block=%p\n", nlmdbg_cookie2a(cookie), block);
196 kref_get(&block->b_count);
197 return block;
198}
199
200/*
201 * Create a block and initialize it.
202 *
203 * Note: we explicitly set the cookie of the grant reply to that of
204 * the blocked lock request. The spec explicitly mentions that the client
205 * should _not_ rely on the callback containing the same cookie as the
206 * request, but (as I found out later) that's because some implementations
207 * do just this. Never mind the standards comittees, they support our
208 * logging industries.
209 *
210 * 10 years later: I hope we can safely ignore these old and broken
211 * clients by now. Let's fix this so we can uniquely identify an incoming
212 * GRANTED_RES message by cookie, without having to rely on the client's IP
213 * address. --okir
214 */
215static struct nlm_block *
216nlmsvc_create_block(struct svc_rqst *rqstp, struct nlm_host *host,
217 struct nlm_file *file, struct nlm_lock *lock,
218 struct nlm_cookie *cookie)
219{
220 struct nlm_block *block;
221 struct nlm_rqst *call = NULL;
222
223 call = nlm_alloc_call(host);
224 if (call == NULL)
225 return NULL;
226
227 /* Allocate memory for block, and initialize arguments */
228 block = kzalloc(sizeof(*block), GFP_KERNEL);
229 if (block == NULL)
230 goto failed;
231 kref_init(&block->b_count);
232 INIT_LIST_HEAD(&block->b_list);
233 INIT_LIST_HEAD(&block->b_flist);
234
235 if (!nlmsvc_setgrantargs(call, lock))
236 goto failed_free;
237
238 /* Set notifier function for VFS, and init args */
239 call->a_args.lock.fl.fl_flags |= FL_SLEEP;
240 call->a_args.lock.fl.fl_lmops = &nlmsvc_lock_operations;
241 nlmclnt_next_cookie(&call->a_args.cookie);
242
243 dprintk("lockd: created block %p...\n", block);
244
245 /* Create and initialize the block */
246 block->b_daemon = rqstp->rq_server;
247 block->b_host = host;
248 block->b_file = file;
249 file->f_count++;
250
251 /* Add to file's list of blocks */
252 list_add(&block->b_flist, &file->f_blocks);
253
254 /* Set up RPC arguments for callback */
255 block->b_call = call;
256 call->a_flags = RPC_TASK_ASYNC;
257 call->a_block = block;
258
259 return block;
260
261failed_free:
262 kfree(block);
263failed:
264 nlmsvc_release_call(call);
265 return NULL;
266}
267
268/*
269 * Delete a block.
270 * It is the caller's responsibility to check whether the file
271 * can be closed hereafter.
272 */
273static int nlmsvc_unlink_block(struct nlm_block *block)
274{
275 int status;
276 dprintk("lockd: unlinking block %p...\n", block);
277
278 /* Remove block from list */
279 status = posix_unblock_lock(&block->b_call->a_args.lock.fl);
280 nlmsvc_remove_block(block);
281 return status;
282}
283
284static void nlmsvc_free_block(struct kref *kref)
285{
286 struct nlm_block *block = container_of(kref, struct nlm_block, b_count);
287 struct nlm_file *file = block->b_file;
288
289 dprintk("lockd: freeing block %p...\n", block);
290
291 /* Remove block from file's list of blocks */
292 list_del_init(&block->b_flist);
293 mutex_unlock(&file->f_mutex);
294
295 nlmsvc_freegrantargs(block->b_call);
296 nlmsvc_release_call(block->b_call);
297 nlm_release_file(block->b_file);
298 kfree(block);
299}
300
301static void nlmsvc_release_block(struct nlm_block *block)
302{
303 if (block != NULL)
304 kref_put_mutex(&block->b_count, nlmsvc_free_block, &block->b_file->f_mutex);
305}
306
307/*
308 * Loop over all blocks and delete blocks held by
309 * a matching host.
310 */
311void nlmsvc_traverse_blocks(struct nlm_host *host,
312 struct nlm_file *file,
313 nlm_host_match_fn_t match)
314{
315 struct nlm_block *block, *next;
316
317restart:
318 mutex_lock(&file->f_mutex);
319 list_for_each_entry_safe(block, next, &file->f_blocks, b_flist) {
320 if (!match(block->b_host, host))
321 continue;
322 /* Do not destroy blocks that are not on
323 * the global retry list - why? */
324 if (list_empty(&block->b_list))
325 continue;
326 kref_get(&block->b_count);
327 mutex_unlock(&file->f_mutex);
328 nlmsvc_unlink_block(block);
329 nlmsvc_release_block(block);
330 goto restart;
331 }
332 mutex_unlock(&file->f_mutex);
333}
334
335/*
336 * Initialize arguments for GRANTED call. The nlm_rqst structure
337 * has been cleared already.
338 */
339static int nlmsvc_setgrantargs(struct nlm_rqst *call, struct nlm_lock *lock)
340{
341 locks_copy_lock(&call->a_args.lock.fl, &lock->fl);
342 memcpy(&call->a_args.lock.fh, &lock->fh, sizeof(call->a_args.lock.fh));
343 call->a_args.lock.caller = utsname()->nodename;
344 call->a_args.lock.oh.len = lock->oh.len;
345
346 /* set default data area */
347 call->a_args.lock.oh.data = call->a_owner;
348 call->a_args.lock.svid = lock->fl.fl_pid;
349
350 if (lock->oh.len > NLMCLNT_OHSIZE) {
351 void *data = kmalloc(lock->oh.len, GFP_KERNEL);
352 if (!data)
353 return 0;
354 call->a_args.lock.oh.data = (u8 *) data;
355 }
356
357 memcpy(call->a_args.lock.oh.data, lock->oh.data, lock->oh.len);
358 return 1;
359}
360
361static void nlmsvc_freegrantargs(struct nlm_rqst *call)
362{
363 if (call->a_args.lock.oh.data != call->a_owner)
364 kfree(call->a_args.lock.oh.data);
365
366 locks_release_private(&call->a_args.lock.fl);
367}
368
369/*
370 * Deferred lock request handling for non-blocking lock
371 */
372static __be32
373nlmsvc_defer_lock_rqst(struct svc_rqst *rqstp, struct nlm_block *block)
374{
375 __be32 status = nlm_lck_denied_nolocks;
376
377 block->b_flags |= B_QUEUED;
378
379 nlmsvc_insert_block(block, NLM_TIMEOUT);
380
381 block->b_cache_req = &rqstp->rq_chandle;
382 if (rqstp->rq_chandle.defer) {
383 block->b_deferred_req =
384 rqstp->rq_chandle.defer(block->b_cache_req);
385 if (block->b_deferred_req != NULL)
386 status = nlm_drop_reply;
387 }
388 dprintk("lockd: nlmsvc_defer_lock_rqst block %p flags %d status %d\n",
389 block, block->b_flags, ntohl(status));
390
391 return status;
392}
393
394/*
395 * Attempt to establish a lock, and if it can't be granted, block it
396 * if required.
397 */
398__be32
399nlmsvc_lock(struct svc_rqst *rqstp, struct nlm_file *file,
400 struct nlm_host *host, struct nlm_lock *lock, int wait,
401 struct nlm_cookie *cookie, int reclaim)
402{
403 struct nlm_block *block = NULL;
404 int error;
405 __be32 ret;
406
407 dprintk("lockd: nlmsvc_lock(%s/%ld, ty=%d, pi=%d, %Ld-%Ld, bl=%d)\n",
408 file_inode(file->f_file)->i_sb->s_id,
409 file_inode(file->f_file)->i_ino,
410 lock->fl.fl_type, lock->fl.fl_pid,
411 (long long)lock->fl.fl_start,
412 (long long)lock->fl.fl_end,
413 wait);
414
415 /* Lock file against concurrent access */
416 mutex_lock(&file->f_mutex);
417 /* Get existing block (in case client is busy-waiting)
418 * or create new block
419 */
420 block = nlmsvc_lookup_block(file, lock);
421 if (block == NULL) {
422 block = nlmsvc_create_block(rqstp, host, file, lock, cookie);
423 ret = nlm_lck_denied_nolocks;
424 if (block == NULL)
425 goto out;
426 lock = &block->b_call->a_args.lock;
427 } else
428 lock->fl.fl_flags &= ~FL_SLEEP;
429
430 if (block->b_flags & B_QUEUED) {
431 dprintk("lockd: nlmsvc_lock deferred block %p flags %d\n",
432 block, block->b_flags);
433 if (block->b_granted) {
434 nlmsvc_unlink_block(block);
435 ret = nlm_granted;
436 goto out;
437 }
438 if (block->b_flags & B_TIMED_OUT) {
439 nlmsvc_unlink_block(block);
440 ret = nlm_lck_denied;
441 goto out;
442 }
443 ret = nlm_drop_reply;
444 goto out;
445 }
446
447 if (locks_in_grace(SVC_NET(rqstp)) && !reclaim) {
448 ret = nlm_lck_denied_grace_period;
449 goto out;
450 }
451 if (reclaim && !locks_in_grace(SVC_NET(rqstp))) {
452 ret = nlm_lck_denied_grace_period;
453 goto out;
454 }
455
456 if (!wait)
457 lock->fl.fl_flags &= ~FL_SLEEP;
458 error = vfs_lock_file(file->f_file, F_SETLK, &lock->fl, NULL);
459 lock->fl.fl_flags &= ~FL_SLEEP;
460
461 dprintk("lockd: vfs_lock_file returned %d\n", error);
462 switch (error) {
463 case 0:
464 ret = nlm_granted;
465 goto out;
466 case -EAGAIN:
467 /*
468 * If this is a blocking request for an
469 * already pending lock request then we need
470 * to put it back on lockd's block list
471 */
472 if (wait)
473 break;
474 ret = nlm_lck_denied;
475 goto out;
476 case FILE_LOCK_DEFERRED:
477 if (wait)
478 break;
479 /* Filesystem lock operation is in progress
480 Add it to the queue waiting for callback */
481 ret = nlmsvc_defer_lock_rqst(rqstp, block);
482 goto out;
483 case -EDEADLK:
484 ret = nlm_deadlock;
485 goto out;
486 default: /* includes ENOLCK */
487 ret = nlm_lck_denied_nolocks;
488 goto out;
489 }
490
491 ret = nlm_lck_blocked;
492
493 /* Append to list of blocked */
494 nlmsvc_insert_block(block, NLM_NEVER);
495out:
496 mutex_unlock(&file->f_mutex);
497 nlmsvc_release_block(block);
498 dprintk("lockd: nlmsvc_lock returned %u\n", ret);
499 return ret;
500}
501
502/*
503 * Test for presence of a conflicting lock.
504 */
505__be32
506nlmsvc_testlock(struct svc_rqst *rqstp, struct nlm_file *file,
507 struct nlm_host *host, struct nlm_lock *lock,
508 struct nlm_lock *conflock, struct nlm_cookie *cookie)
509{
510 int error;
511 __be32 ret;
512
513 dprintk("lockd: nlmsvc_testlock(%s/%ld, ty=%d, %Ld-%Ld)\n",
514 file_inode(file->f_file)->i_sb->s_id,
515 file_inode(file->f_file)->i_ino,
516 lock->fl.fl_type,
517 (long long)lock->fl.fl_start,
518 (long long)lock->fl.fl_end);
519
520 if (locks_in_grace(SVC_NET(rqstp))) {
521 ret = nlm_lck_denied_grace_period;
522 goto out;
523 }
524
525 error = vfs_test_lock(file->f_file, &lock->fl);
526 if (error) {
527 /* We can't currently deal with deferred test requests */
528 if (error == FILE_LOCK_DEFERRED)
529 WARN_ON_ONCE(1);
530
531 ret = nlm_lck_denied_nolocks;
532 goto out;
533 }
534
535 if (lock->fl.fl_type == F_UNLCK) {
536 ret = nlm_granted;
537 goto out;
538 }
539
540 dprintk("lockd: conflicting lock(ty=%d, %Ld-%Ld)\n",
541 lock->fl.fl_type, (long long)lock->fl.fl_start,
542 (long long)lock->fl.fl_end);
543 conflock->caller = "somehost"; /* FIXME */
544 conflock->len = strlen(conflock->caller);
545 conflock->oh.len = 0; /* don't return OH info */
546 conflock->svid = lock->fl.fl_pid;
547 conflock->fl.fl_type = lock->fl.fl_type;
548 conflock->fl.fl_start = lock->fl.fl_start;
549 conflock->fl.fl_end = lock->fl.fl_end;
550 locks_release_private(&lock->fl);
551 ret = nlm_lck_denied;
552out:
553 return ret;
554}
555
556/*
557 * Remove a lock.
558 * This implies a CANCEL call: We send a GRANT_MSG, the client replies
559 * with a GRANT_RES call which gets lost, and calls UNLOCK immediately
560 * afterwards. In this case the block will still be there, and hence
561 * must be removed.
562 */
563__be32
564nlmsvc_unlock(struct net *net, struct nlm_file *file, struct nlm_lock *lock)
565{
566 int error;
567
568 dprintk("lockd: nlmsvc_unlock(%s/%ld, pi=%d, %Ld-%Ld)\n",
569 file_inode(file->f_file)->i_sb->s_id,
570 file_inode(file->f_file)->i_ino,
571 lock->fl.fl_pid,
572 (long long)lock->fl.fl_start,
573 (long long)lock->fl.fl_end);
574
575 /* First, cancel any lock that might be there */
576 nlmsvc_cancel_blocked(net, file, lock);
577
578 lock->fl.fl_type = F_UNLCK;
579 error = vfs_lock_file(file->f_file, F_SETLK, &lock->fl, NULL);
580
581 return (error < 0)? nlm_lck_denied_nolocks : nlm_granted;
582}
583
584/*
585 * Cancel a previously blocked request.
586 *
587 * A cancel request always overrides any grant that may currently
588 * be in progress.
589 * The calling procedure must check whether the file can be closed.
590 */
591__be32
592nlmsvc_cancel_blocked(struct net *net, struct nlm_file *file, struct nlm_lock *lock)
593{
594 struct nlm_block *block;
595 int status = 0;
596
597 dprintk("lockd: nlmsvc_cancel(%s/%ld, pi=%d, %Ld-%Ld)\n",
598 file_inode(file->f_file)->i_sb->s_id,
599 file_inode(file->f_file)->i_ino,
600 lock->fl.fl_pid,
601 (long long)lock->fl.fl_start,
602 (long long)lock->fl.fl_end);
603
604 if (locks_in_grace(net))
605 return nlm_lck_denied_grace_period;
606
607 mutex_lock(&file->f_mutex);
608 block = nlmsvc_lookup_block(file, lock);
609 mutex_unlock(&file->f_mutex);
610 if (block != NULL) {
611 vfs_cancel_lock(block->b_file->f_file,
612 &block->b_call->a_args.lock.fl);
613 status = nlmsvc_unlink_block(block);
614 nlmsvc_release_block(block);
615 }
616 return status ? nlm_lck_denied : nlm_granted;
617}
618
619/*
620 * This is a callback from the filesystem for VFS file lock requests.
621 * It will be used if lm_grant is defined and the filesystem can not
622 * respond to the request immediately.
623 * For SETLK or SETLKW request it will get the local posix lock.
624 * In all cases it will move the block to the head of nlm_blocked q where
625 * nlmsvc_retry_blocked() can send back a reply for SETLKW or revisit the
626 * deferred rpc for GETLK and SETLK.
627 */
628static void
629nlmsvc_update_deferred_block(struct nlm_block *block, int result)
630{
631 block->b_flags |= B_GOT_CALLBACK;
632 if (result == 0)
633 block->b_granted = 1;
634 else
635 block->b_flags |= B_TIMED_OUT;
636}
637
638static int nlmsvc_grant_deferred(struct file_lock *fl, int result)
639{
640 struct nlm_block *block;
641 int rc = -ENOENT;
642
643 spin_lock(&nlm_blocked_lock);
644 list_for_each_entry(block, &nlm_blocked, b_list) {
645 if (nlm_compare_locks(&block->b_call->a_args.lock.fl, fl)) {
646 dprintk("lockd: nlmsvc_notify_blocked block %p flags %d\n",
647 block, block->b_flags);
648 if (block->b_flags & B_QUEUED) {
649 if (block->b_flags & B_TIMED_OUT) {
650 rc = -ENOLCK;
651 break;
652 }
653 nlmsvc_update_deferred_block(block, result);
654 } else if (result == 0)
655 block->b_granted = 1;
656
657 nlmsvc_insert_block_locked(block, 0);
658 svc_wake_up(block->b_daemon);
659 rc = 0;
660 break;
661 }
662 }
663 spin_unlock(&nlm_blocked_lock);
664 if (rc == -ENOENT)
665 printk(KERN_WARNING "lockd: grant for unknown block\n");
666 return rc;
667}
668
669/*
670 * Unblock a blocked lock request. This is a callback invoked from the
671 * VFS layer when a lock on which we blocked is removed.
672 *
673 * This function doesn't grant the blocked lock instantly, but rather moves
674 * the block to the head of nlm_blocked where it can be picked up by lockd.
675 */
676static void
677nlmsvc_notify_blocked(struct file_lock *fl)
678{
679 struct nlm_block *block;
680
681 dprintk("lockd: VFS unblock notification for block %p\n", fl);
682 spin_lock(&nlm_blocked_lock);
683 list_for_each_entry(block, &nlm_blocked, b_list) {
684 if (nlm_compare_locks(&block->b_call->a_args.lock.fl, fl)) {
685 nlmsvc_insert_block_locked(block, 0);
686 spin_unlock(&nlm_blocked_lock);
687 svc_wake_up(block->b_daemon);
688 return;
689 }
690 }
691 spin_unlock(&nlm_blocked_lock);
692 printk(KERN_WARNING "lockd: notification for unknown block!\n");
693}
694
695static int nlmsvc_same_owner(struct file_lock *fl1, struct file_lock *fl2)
696{
697 return fl1->fl_owner == fl2->fl_owner && fl1->fl_pid == fl2->fl_pid;
698}
699
700/*
701 * Since NLM uses two "keys" for tracking locks, we need to hash them down
702 * to one for the blocked_hash. Here, we're just xor'ing the host address
703 * with the pid in order to create a key value for picking a hash bucket.
704 */
705static unsigned long
706nlmsvc_owner_key(struct file_lock *fl)
707{
708 return (unsigned long)fl->fl_owner ^ (unsigned long)fl->fl_pid;
709}
710
711const struct lock_manager_operations nlmsvc_lock_operations = {
712 .lm_compare_owner = nlmsvc_same_owner,
713 .lm_owner_key = nlmsvc_owner_key,
714 .lm_notify = nlmsvc_notify_blocked,
715 .lm_grant = nlmsvc_grant_deferred,
716};
717
718/*
719 * Try to claim a lock that was previously blocked.
720 *
721 * Note that we use both the RPC_GRANTED_MSG call _and_ an async
722 * RPC thread when notifying the client. This seems like overkill...
723 * Here's why:
724 * - we don't want to use a synchronous RPC thread, otherwise
725 * we might find ourselves hanging on a dead portmapper.
726 * - Some lockd implementations (e.g. HP) don't react to
727 * RPC_GRANTED calls; they seem to insist on RPC_GRANTED_MSG calls.
728 */
729static void
730nlmsvc_grant_blocked(struct nlm_block *block)
731{
732 struct nlm_file *file = block->b_file;
733 struct nlm_lock *lock = &block->b_call->a_args.lock;
734 int error;
735 loff_t fl_start, fl_end;
736
737 dprintk("lockd: grant blocked lock %p\n", block);
738
739 kref_get(&block->b_count);
740
741 /* Unlink block request from list */
742 nlmsvc_unlink_block(block);
743
744 /* If b_granted is true this means we've been here before.
745 * Just retry the grant callback, possibly refreshing the RPC
746 * binding */
747 if (block->b_granted) {
748 nlm_rebind_host(block->b_host);
749 goto callback;
750 }
751
752 /* Try the lock operation again */
753 /* vfs_lock_file() can mangle fl_start and fl_end, but we need
754 * them unchanged for the GRANT_MSG
755 */
756 lock->fl.fl_flags |= FL_SLEEP;
757 fl_start = lock->fl.fl_start;
758 fl_end = lock->fl.fl_end;
759 error = vfs_lock_file(file->f_file, F_SETLK, &lock->fl, NULL);
760 lock->fl.fl_flags &= ~FL_SLEEP;
761 lock->fl.fl_start = fl_start;
762 lock->fl.fl_end = fl_end;
763
764 switch (error) {
765 case 0:
766 break;
767 case FILE_LOCK_DEFERRED:
768 dprintk("lockd: lock still blocked error %d\n", error);
769 nlmsvc_insert_block(block, NLM_NEVER);
770 nlmsvc_release_block(block);
771 return;
772 default:
773 printk(KERN_WARNING "lockd: unexpected error %d in %s!\n",
774 -error, __func__);
775 nlmsvc_insert_block(block, 10 * HZ);
776 nlmsvc_release_block(block);
777 return;
778 }
779
780callback:
781 /* Lock was granted by VFS. */
782 dprintk("lockd: GRANTing blocked lock.\n");
783 block->b_granted = 1;
784
785 /* keep block on the list, but don't reattempt until the RPC
786 * completes or the submission fails
787 */
788 nlmsvc_insert_block(block, NLM_NEVER);
789
790 /* Call the client -- use a soft RPC task since nlmsvc_retry_blocked
791 * will queue up a new one if this one times out
792 */
793 error = nlm_async_call(block->b_call, NLMPROC_GRANTED_MSG,
794 &nlmsvc_grant_ops);
795
796 /* RPC submission failed, wait a bit and retry */
797 if (error < 0)
798 nlmsvc_insert_block(block, 10 * HZ);
799}
800
801/*
802 * This is the callback from the RPC layer when the NLM_GRANTED_MSG
803 * RPC call has succeeded or timed out.
804 * Like all RPC callbacks, it is invoked by the rpciod process, so it
805 * better not sleep. Therefore, we put the blocked lock on the nlm_blocked
806 * chain once more in order to have it removed by lockd itself (which can
807 * then sleep on the file semaphore without disrupting e.g. the nfs client).
808 */
809static void nlmsvc_grant_callback(struct rpc_task *task, void *data)
810{
811 struct nlm_rqst *call = data;
812 struct nlm_block *block = call->a_block;
813 unsigned long timeout;
814
815 dprintk("lockd: GRANT_MSG RPC callback\n");
816
817 spin_lock(&nlm_blocked_lock);
818 /* if the block is not on a list at this point then it has
819 * been invalidated. Don't try to requeue it.
820 *
821 * FIXME: it's possible that the block is removed from the list
822 * after this check but before the nlmsvc_insert_block. In that
823 * case it will be added back. Perhaps we need better locking
824 * for nlm_blocked?
825 */
826 if (list_empty(&block->b_list))
827 goto out;
828
829 /* Technically, we should down the file semaphore here. Since we
830 * move the block towards the head of the queue only, no harm
831 * can be done, though. */
832 if (task->tk_status < 0) {
833 /* RPC error: Re-insert for retransmission */
834 timeout = 10 * HZ;
835 } else {
836 /* Call was successful, now wait for client callback */
837 timeout = 60 * HZ;
838 }
839 nlmsvc_insert_block_locked(block, timeout);
840 svc_wake_up(block->b_daemon);
841out:
842 spin_unlock(&nlm_blocked_lock);
843}
844
845/*
846 * FIXME: nlmsvc_release_block() grabs a mutex. This is not allowed for an
847 * .rpc_release rpc_call_op
848 */
849static void nlmsvc_grant_release(void *data)
850{
851 struct nlm_rqst *call = data;
852 nlmsvc_release_block(call->a_block);
853}
854
855static const struct rpc_call_ops nlmsvc_grant_ops = {
856 .rpc_call_done = nlmsvc_grant_callback,
857 .rpc_release = nlmsvc_grant_release,
858};
859
860/*
861 * We received a GRANT_RES callback. Try to find the corresponding
862 * block.
863 */
864void
865nlmsvc_grant_reply(struct nlm_cookie *cookie, __be32 status)
866{
867 struct nlm_block *block;
868
869 dprintk("grant_reply: looking for cookie %x, s=%d \n",
870 *(unsigned int *)(cookie->data), status);
871 if (!(block = nlmsvc_find_block(cookie)))
872 return;
873
874 if (status == nlm_lck_denied_grace_period) {
875 /* Try again in a couple of seconds */
876 nlmsvc_insert_block(block, 10 * HZ);
877 } else {
878 /*
879 * Lock is now held by client, or has been rejected.
880 * In both cases, the block should be removed.
881 */
882 nlmsvc_unlink_block(block);
883 }
884 nlmsvc_release_block(block);
885}
886
887/* Helper function to handle retry of a deferred block.
888 * If it is a blocking lock, call grant_blocked.
889 * For a non-blocking lock or test lock, revisit the request.
890 */
891static void
892retry_deferred_block(struct nlm_block *block)
893{
894 if (!(block->b_flags & B_GOT_CALLBACK))
895 block->b_flags |= B_TIMED_OUT;
896 nlmsvc_insert_block(block, NLM_TIMEOUT);
897 dprintk("revisit block %p flags %d\n", block, block->b_flags);
898 if (block->b_deferred_req) {
899 block->b_deferred_req->revisit(block->b_deferred_req, 0);
900 block->b_deferred_req = NULL;
901 }
902}
903
904/*
905 * Retry all blocked locks that have been notified. This is where lockd
906 * picks up locks that can be granted, or grant notifications that must
907 * be retransmitted.
908 */
909unsigned long
910nlmsvc_retry_blocked(void)
911{
912 unsigned long timeout = MAX_SCHEDULE_TIMEOUT;
913 struct nlm_block *block;
914
915 spin_lock(&nlm_blocked_lock);
916 while (!list_empty(&nlm_blocked) && !kthread_should_stop()) {
917 block = list_entry(nlm_blocked.next, struct nlm_block, b_list);
918
919 if (block->b_when == NLM_NEVER)
920 break;
921 if (time_after(block->b_when, jiffies)) {
922 timeout = block->b_when - jiffies;
923 break;
924 }
925 spin_unlock(&nlm_blocked_lock);
926
927 dprintk("nlmsvc_retry_blocked(%p, when=%ld)\n",
928 block, block->b_when);
929 if (block->b_flags & B_QUEUED) {
930 dprintk("nlmsvc_retry_blocked delete block (%p, granted=%d, flags=%d)\n",
931 block, block->b_granted, block->b_flags);
932 retry_deferred_block(block);
933 } else
934 nlmsvc_grant_blocked(block);
935 spin_lock(&nlm_blocked_lock);
936 }
937 spin_unlock(&nlm_blocked_lock);
938
939 return timeout;
940}