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_xprt.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;
49
50/*
51 * The list of blocked locks to retry
52 */
53static LIST_HEAD(nlm_blocked);
54static DEFINE_SPINLOCK(nlm_blocked_lock);
55
56#if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
57static const char *nlmdbg_cookie2a(const struct nlm_cookie *cookie)
58{
59 /*
60 * We can get away with a static buffer because this is only called
61 * from lockd, which is single-threaded.
62 */
63 static char buf[2*NLM_MAXCOOKIELEN+1];
64 unsigned int i, len = sizeof(buf);
65 char *p = buf;
66
67 len--; /* allow for trailing \0 */
68 if (len < 3)
69 return "???";
70 for (i = 0 ; i < cookie->len ; i++) {
71 if (len < 2) {
72 strcpy(p-3, "...");
73 break;
74 }
75 sprintf(p, "%02x", cookie->data[i]);
76 p += 2;
77 len -= 2;
78 }
79 *p = '\0';
80
81 return buf;
82}
83#endif
84
85/*
86 * Insert a blocked lock into the global list
87 */
88static void
89nlmsvc_insert_block_locked(struct nlm_block *block, unsigned long when)
90{
91 struct nlm_block *b;
92 struct list_head *pos;
93
94 dprintk("lockd: nlmsvc_insert_block(%p, %ld)\n", block, when);
95 if (list_empty(&block->b_list)) {
96 kref_get(&block->b_count);
97 } else {
98 list_del_init(&block->b_list);
99 }
100
101 pos = &nlm_blocked;
102 if (when != NLM_NEVER) {
103 if ((when += jiffies) == NLM_NEVER)
104 when ++;
105 list_for_each(pos, &nlm_blocked) {
106 b = list_entry(pos, struct nlm_block, b_list);
107 if (time_after(b->b_when,when) || b->b_when == NLM_NEVER)
108 break;
109 }
110 /* On normal exit from the loop, pos == &nlm_blocked,
111 * so we will be adding to the end of the list - good
112 */
113 }
114
115 list_add_tail(&block->b_list, pos);
116 block->b_when = when;
117}
118
119static void nlmsvc_insert_block(struct nlm_block *block, unsigned long when)
120{
121 spin_lock(&nlm_blocked_lock);
122 nlmsvc_insert_block_locked(block, when);
123 spin_unlock(&nlm_blocked_lock);
124}
125
126/*
127 * Remove a block from the global list
128 */
129static inline void
130nlmsvc_remove_block(struct nlm_block *block)
131{
132 if (!list_empty(&block->b_list)) {
133 spin_lock(&nlm_blocked_lock);
134 list_del_init(&block->b_list);
135 spin_unlock(&nlm_blocked_lock);
136 nlmsvc_release_block(block);
137 }
138}
139
140/*
141 * Find a block for a given lock
142 */
143static struct nlm_block *
144nlmsvc_lookup_block(struct nlm_file *file, struct nlm_lock *lock)
145{
146 struct nlm_block *block;
147 struct file_lock *fl;
148
149 dprintk("lockd: nlmsvc_lookup_block f=%p pd=%d %Ld-%Ld ty=%d\n",
150 file, lock->fl.fl_pid,
151 (long long)lock->fl.fl_start,
152 (long long)lock->fl.fl_end, lock->fl.fl_type);
153 list_for_each_entry(block, &nlm_blocked, b_list) {
154 fl = &block->b_call->a_args.lock.fl;
155 dprintk("lockd: check f=%p pd=%d %Ld-%Ld ty=%d cookie=%s\n",
156 block->b_file, fl->fl_pid,
157 (long long)fl->fl_start,
158 (long long)fl->fl_end, fl->fl_type,
159 nlmdbg_cookie2a(&block->b_call->a_args.cookie));
160 if (block->b_file == file && nlm_compare_locks(fl, &lock->fl)) {
161 kref_get(&block->b_count);
162 return block;
163 }
164 }
165
166 return NULL;
167}
168
169static inline int nlm_cookie_match(struct nlm_cookie *a, struct nlm_cookie *b)
170{
171 if (a->len != b->len)
172 return 0;
173 if (memcmp(a->data, b->data, a->len))
174 return 0;
175 return 1;
176}
177
178/*
179 * Find a block with a given NLM cookie.
180 */
181static inline struct nlm_block *
182nlmsvc_find_block(struct nlm_cookie *cookie)
183{
184 struct nlm_block *block;
185
186 list_for_each_entry(block, &nlm_blocked, b_list) {
187 if (nlm_cookie_match(&block->b_call->a_args.cookie,cookie))
188 goto found;
189 }
190
191 return NULL;
192
193found:
194 dprintk("nlmsvc_find_block(%s): block=%p\n", nlmdbg_cookie2a(cookie), block);
195 kref_get(&block->b_count);
196 return block;
197}
198
199/*
200 * Create a block and initialize it.
201 *
202 * Note: we explicitly set the cookie of the grant reply to that of
203 * the blocked lock request. The spec explicitly mentions that the client
204 * should _not_ rely on the callback containing the same cookie as the
205 * request, but (as I found out later) that's because some implementations
206 * do just this. Never mind the standards comittees, they support our
207 * logging industries.
208 *
209 * 10 years later: I hope we can safely ignore these old and broken
210 * clients by now. Let's fix this so we can uniquely identify an incoming
211 * GRANTED_RES message by cookie, without having to rely on the client's IP
212 * address. --okir
213 */
214static struct nlm_block *
215nlmsvc_create_block(struct svc_rqst *rqstp, struct nlm_host *host,
216 struct nlm_file *file, struct nlm_lock *lock,
217 struct nlm_cookie *cookie)
218{
219 struct nlm_block *block;
220 struct nlm_rqst *call = NULL;
221
222 call = nlm_alloc_call(host);
223 if (call == NULL)
224 return NULL;
225
226 /* Allocate memory for block, and initialize arguments */
227 block = kzalloc(sizeof(*block), GFP_KERNEL);
228 if (block == NULL)
229 goto failed;
230 kref_init(&block->b_count);
231 INIT_LIST_HEAD(&block->b_list);
232 INIT_LIST_HEAD(&block->b_flist);
233
234 if (!nlmsvc_setgrantargs(call, lock))
235 goto failed_free;
236
237 /* Set notifier function for VFS, and init args */
238 call->a_args.lock.fl.fl_flags |= FL_SLEEP;
239 call->a_args.lock.fl.fl_lmops = &nlmsvc_lock_operations;
240 nlmclnt_next_cookie(&call->a_args.cookie);
241
242 dprintk("lockd: created block %p...\n", block);
243
244 /* Create and initialize the block */
245 block->b_daemon = rqstp->rq_server;
246 block->b_host = host;
247 block->b_file = file;
248 file->f_count++;
249
250 /* Add to file's list of blocks */
251 list_add(&block->b_flist, &file->f_blocks);
252
253 /* Set up RPC arguments for callback */
254 block->b_call = call;
255 call->a_flags = RPC_TASK_ASYNC;
256 call->a_block = block;
257
258 return block;
259
260failed_free:
261 kfree(block);
262failed:
263 nlmsvc_release_call(call);
264 return NULL;
265}
266
267/*
268 * Delete a block.
269 * It is the caller's responsibility to check whether the file
270 * can be closed hereafter.
271 */
272static int nlmsvc_unlink_block(struct nlm_block *block)
273{
274 int status;
275 dprintk("lockd: unlinking block %p...\n", block);
276
277 /* Remove block from list */
278 status = posix_unblock_lock(&block->b_call->a_args.lock.fl);
279 nlmsvc_remove_block(block);
280 return status;
281}
282
283static void nlmsvc_free_block(struct kref *kref)
284{
285 struct nlm_block *block = container_of(kref, struct nlm_block, b_count);
286 struct nlm_file *file = block->b_file;
287
288 dprintk("lockd: freeing block %p...\n", block);
289
290 /* Remove block from file's list of blocks */
291 list_del_init(&block->b_flist);
292 mutex_unlock(&file->f_mutex);
293
294 nlmsvc_freegrantargs(block->b_call);
295 nlmsvc_release_call(block->b_call);
296 nlm_release_file(block->b_file);
297 kfree(block);
298}
299
300static void nlmsvc_release_block(struct nlm_block *block)
301{
302 if (block != NULL)
303 kref_put_mutex(&block->b_count, nlmsvc_free_block, &block->b_file->f_mutex);
304}
305
306/*
307 * Loop over all blocks and delete blocks held by
308 * a matching host.
309 */
310void nlmsvc_traverse_blocks(struct nlm_host *host,
311 struct nlm_file *file,
312 nlm_host_match_fn_t match)
313{
314 struct nlm_block *block, *next;
315
316restart:
317 mutex_lock(&file->f_mutex);
318 list_for_each_entry_safe(block, next, &file->f_blocks, b_flist) {
319 if (!match(block->b_host, host))
320 continue;
321 /* Do not destroy blocks that are not on
322 * the global retry list - why? */
323 if (list_empty(&block->b_list))
324 continue;
325 kref_get(&block->b_count);
326 mutex_unlock(&file->f_mutex);
327 nlmsvc_unlink_block(block);
328 nlmsvc_release_block(block);
329 goto restart;
330 }
331 mutex_unlock(&file->f_mutex);
332}
333
334/*
335 * Initialize arguments for GRANTED call. The nlm_rqst structure
336 * has been cleared already.
337 */
338static int nlmsvc_setgrantargs(struct nlm_rqst *call, struct nlm_lock *lock)
339{
340 locks_copy_lock(&call->a_args.lock.fl, &lock->fl);
341 memcpy(&call->a_args.lock.fh, &lock->fh, sizeof(call->a_args.lock.fh));
342 call->a_args.lock.caller = utsname()->nodename;
343 call->a_args.lock.oh.len = lock->oh.len;
344
345 /* set default data area */
346 call->a_args.lock.oh.data = call->a_owner;
347 call->a_args.lock.svid = lock->fl.fl_pid;
348
349 if (lock->oh.len > NLMCLNT_OHSIZE) {
350 void *data = kmalloc(lock->oh.len, GFP_KERNEL);
351 if (!data)
352 return 0;
353 call->a_args.lock.oh.data = (u8 *) data;
354 }
355
356 memcpy(call->a_args.lock.oh.data, lock->oh.data, lock->oh.len);
357 return 1;
358}
359
360static void nlmsvc_freegrantargs(struct nlm_rqst *call)
361{
362 if (call->a_args.lock.oh.data != call->a_owner)
363 kfree(call->a_args.lock.oh.data);
364
365 locks_release_private(&call->a_args.lock.fl);
366}
367
368/*
369 * Deferred lock request handling for non-blocking lock
370 */
371static __be32
372nlmsvc_defer_lock_rqst(struct svc_rqst *rqstp, struct nlm_block *block)
373{
374 __be32 status = nlm_lck_denied_nolocks;
375
376 block->b_flags |= B_QUEUED;
377
378 nlmsvc_insert_block(block, NLM_TIMEOUT);
379
380 block->b_cache_req = &rqstp->rq_chandle;
381 if (rqstp->rq_chandle.defer) {
382 block->b_deferred_req =
383 rqstp->rq_chandle.defer(block->b_cache_req);
384 if (block->b_deferred_req != NULL)
385 status = nlm_drop_reply;
386 }
387 dprintk("lockd: nlmsvc_defer_lock_rqst block %p flags %d status %d\n",
388 block, block->b_flags, ntohl(status));
389
390 return status;
391}
392
393/*
394 * Attempt to establish a lock, and if it can't be granted, block it
395 * if required.
396 */
397__be32
398nlmsvc_lock(struct svc_rqst *rqstp, struct nlm_file *file,
399 struct nlm_host *host, struct nlm_lock *lock, int wait,
400 struct nlm_cookie *cookie, int reclaim)
401{
402 struct nlm_block *block = NULL;
403 int error;
404 __be32 ret;
405
406 dprintk("lockd: nlmsvc_lock(%s/%ld, ty=%d, pi=%d, %Ld-%Ld, bl=%d)\n",
407 file_inode(file->f_file)->i_sb->s_id,
408 file_inode(file->f_file)->i_ino,
409 lock->fl.fl_type, lock->fl.fl_pid,
410 (long long)lock->fl.fl_start,
411 (long long)lock->fl.fl_end,
412 wait);
413
414 /* Lock file against concurrent access */
415 mutex_lock(&file->f_mutex);
416 /* Get existing block (in case client is busy-waiting)
417 * or create new block
418 */
419 block = nlmsvc_lookup_block(file, lock);
420 if (block == NULL) {
421 block = nlmsvc_create_block(rqstp, host, file, lock, cookie);
422 ret = nlm_lck_denied_nolocks;
423 if (block == NULL)
424 goto out;
425 lock = &block->b_call->a_args.lock;
426 } else
427 lock->fl.fl_flags &= ~FL_SLEEP;
428
429 if (block->b_flags & B_QUEUED) {
430 dprintk("lockd: nlmsvc_lock deferred block %p flags %d\n",
431 block, block->b_flags);
432 if (block->b_granted) {
433 nlmsvc_unlink_block(block);
434 ret = nlm_granted;
435 goto out;
436 }
437 if (block->b_flags & B_TIMED_OUT) {
438 nlmsvc_unlink_block(block);
439 ret = nlm_lck_denied;
440 goto out;
441 }
442 ret = nlm_drop_reply;
443 goto out;
444 }
445
446 if (locks_in_grace(SVC_NET(rqstp)) && !reclaim) {
447 ret = nlm_lck_denied_grace_period;
448 goto out;
449 }
450 if (reclaim && !locks_in_grace(SVC_NET(rqstp))) {
451 ret = nlm_lck_denied_grace_period;
452 goto out;
453 }
454
455 if (!wait)
456 lock->fl.fl_flags &= ~FL_SLEEP;
457 error = vfs_lock_file(file->f_file, F_SETLK, &lock->fl, NULL);
458 lock->fl.fl_flags &= ~FL_SLEEP;
459
460 dprintk("lockd: vfs_lock_file returned %d\n", error);
461 switch (error) {
462 case 0:
463 ret = nlm_granted;
464 goto out;
465 case -EAGAIN:
466 /*
467 * If this is a blocking request for an
468 * already pending lock request then we need
469 * to put it back on lockd's block list
470 */
471 if (wait)
472 break;
473 ret = nlm_lck_denied;
474 goto out;
475 case FILE_LOCK_DEFERRED:
476 if (wait)
477 break;
478 /* Filesystem lock operation is in progress
479 Add it to the queue waiting for callback */
480 ret = nlmsvc_defer_lock_rqst(rqstp, block);
481 goto out;
482 case -EDEADLK:
483 ret = nlm_deadlock;
484 goto out;
485 default: /* includes ENOLCK */
486 ret = nlm_lck_denied_nolocks;
487 goto out;
488 }
489
490 ret = nlm_lck_blocked;
491
492 /* Append to list of blocked */
493 nlmsvc_insert_block(block, NLM_NEVER);
494out:
495 mutex_unlock(&file->f_mutex);
496 nlmsvc_release_block(block);
497 dprintk("lockd: nlmsvc_lock returned %u\n", ret);
498 return ret;
499}
500
501/*
502 * Test for presence of a conflicting lock.
503 */
504__be32
505nlmsvc_testlock(struct svc_rqst *rqstp, struct nlm_file *file,
506 struct nlm_host *host, struct nlm_lock *lock,
507 struct nlm_lock *conflock, struct nlm_cookie *cookie)
508{
509 int error;
510 __be32 ret;
511
512 dprintk("lockd: nlmsvc_testlock(%s/%ld, ty=%d, %Ld-%Ld)\n",
513 file_inode(file->f_file)->i_sb->s_id,
514 file_inode(file->f_file)->i_ino,
515 lock->fl.fl_type,
516 (long long)lock->fl.fl_start,
517 (long long)lock->fl.fl_end);
518
519 if (locks_in_grace(SVC_NET(rqstp))) {
520 ret = nlm_lck_denied_grace_period;
521 goto out;
522 }
523
524 error = vfs_test_lock(file->f_file, &lock->fl);
525 if (error) {
526 /* We can't currently deal with deferred test requests */
527 if (error == FILE_LOCK_DEFERRED)
528 WARN_ON_ONCE(1);
529
530 ret = nlm_lck_denied_nolocks;
531 goto out;
532 }
533
534 if (lock->fl.fl_type == F_UNLCK) {
535 ret = nlm_granted;
536 goto out;
537 }
538
539 dprintk("lockd: conflicting lock(ty=%d, %Ld-%Ld)\n",
540 lock->fl.fl_type, (long long)lock->fl.fl_start,
541 (long long)lock->fl.fl_end);
542 conflock->caller = "somehost"; /* FIXME */
543 conflock->len = strlen(conflock->caller);
544 conflock->oh.len = 0; /* don't return OH info */
545 conflock->svid = lock->fl.fl_pid;
546 conflock->fl.fl_type = lock->fl.fl_type;
547 conflock->fl.fl_start = lock->fl.fl_start;
548 conflock->fl.fl_end = lock->fl.fl_end;
549 locks_release_private(&lock->fl);
550 ret = nlm_lck_denied;
551out:
552 return ret;
553}
554
555/*
556 * Remove a lock.
557 * This implies a CANCEL call: We send a GRANT_MSG, the client replies
558 * with a GRANT_RES call which gets lost, and calls UNLOCK immediately
559 * afterwards. In this case the block will still be there, and hence
560 * must be removed.
561 */
562__be32
563nlmsvc_unlock(struct net *net, struct nlm_file *file, struct nlm_lock *lock)
564{
565 int error;
566
567 dprintk("lockd: nlmsvc_unlock(%s/%ld, pi=%d, %Ld-%Ld)\n",
568 file_inode(file->f_file)->i_sb->s_id,
569 file_inode(file->f_file)->i_ino,
570 lock->fl.fl_pid,
571 (long long)lock->fl.fl_start,
572 (long long)lock->fl.fl_end);
573
574 /* First, cancel any lock that might be there */
575 nlmsvc_cancel_blocked(net, file, lock);
576
577 lock->fl.fl_type = F_UNLCK;
578 error = vfs_lock_file(file->f_file, F_SETLK, &lock->fl, NULL);
579
580 return (error < 0)? nlm_lck_denied_nolocks : nlm_granted;
581}
582
583/*
584 * Cancel a previously blocked request.
585 *
586 * A cancel request always overrides any grant that may currently
587 * be in progress.
588 * The calling procedure must check whether the file can be closed.
589 */
590__be32
591nlmsvc_cancel_blocked(struct net *net, struct nlm_file *file, struct nlm_lock *lock)
592{
593 struct nlm_block *block;
594 int status = 0;
595
596 dprintk("lockd: nlmsvc_cancel(%s/%ld, pi=%d, %Ld-%Ld)\n",
597 file_inode(file->f_file)->i_sb->s_id,
598 file_inode(file->f_file)->i_ino,
599 lock->fl.fl_pid,
600 (long long)lock->fl.fl_start,
601 (long long)lock->fl.fl_end);
602
603 if (locks_in_grace(net))
604 return nlm_lck_denied_grace_period;
605
606 mutex_lock(&file->f_mutex);
607 block = nlmsvc_lookup_block(file, lock);
608 mutex_unlock(&file->f_mutex);
609 if (block != NULL) {
610 vfs_cancel_lock(block->b_file->f_file,
611 &block->b_call->a_args.lock.fl);
612 status = nlmsvc_unlink_block(block);
613 nlmsvc_release_block(block);
614 }
615 return status ? nlm_lck_denied : nlm_granted;
616}
617
618/*
619 * This is a callback from the filesystem for VFS file lock requests.
620 * It will be used if lm_grant is defined and the filesystem can not
621 * respond to the request immediately.
622 * For SETLK or SETLKW request it will get the local posix lock.
623 * In all cases it will move the block to the head of nlm_blocked q where
624 * nlmsvc_retry_blocked() can send back a reply for SETLKW or revisit the
625 * deferred rpc for GETLK and SETLK.
626 */
627static void
628nlmsvc_update_deferred_block(struct nlm_block *block, int result)
629{
630 block->b_flags |= B_GOT_CALLBACK;
631 if (result == 0)
632 block->b_granted = 1;
633 else
634 block->b_flags |= B_TIMED_OUT;
635}
636
637static int nlmsvc_grant_deferred(struct file_lock *fl, int result)
638{
639 struct nlm_block *block;
640 int rc = -ENOENT;
641
642 spin_lock(&nlm_blocked_lock);
643 list_for_each_entry(block, &nlm_blocked, b_list) {
644 if (nlm_compare_locks(&block->b_call->a_args.lock.fl, fl)) {
645 dprintk("lockd: nlmsvc_notify_blocked block %p flags %d\n",
646 block, block->b_flags);
647 if (block->b_flags & B_QUEUED) {
648 if (block->b_flags & B_TIMED_OUT) {
649 rc = -ENOLCK;
650 break;
651 }
652 nlmsvc_update_deferred_block(block, result);
653 } else if (result == 0)
654 block->b_granted = 1;
655
656 nlmsvc_insert_block_locked(block, 0);
657 svc_wake_up(block->b_daemon);
658 rc = 0;
659 break;
660 }
661 }
662 spin_unlock(&nlm_blocked_lock);
663 if (rc == -ENOENT)
664 printk(KERN_WARNING "lockd: grant for unknown block\n");
665 return rc;
666}
667
668/*
669 * Unblock a blocked lock request. This is a callback invoked from the
670 * VFS layer when a lock on which we blocked is removed.
671 *
672 * This function doesn't grant the blocked lock instantly, but rather moves
673 * the block to the head of nlm_blocked where it can be picked up by lockd.
674 */
675static void
676nlmsvc_notify_blocked(struct file_lock *fl)
677{
678 struct nlm_block *block;
679
680 dprintk("lockd: VFS unblock notification for block %p\n", fl);
681 spin_lock(&nlm_blocked_lock);
682 list_for_each_entry(block, &nlm_blocked, b_list) {
683 if (nlm_compare_locks(&block->b_call->a_args.lock.fl, fl)) {
684 nlmsvc_insert_block_locked(block, 0);
685 spin_unlock(&nlm_blocked_lock);
686 svc_wake_up(block->b_daemon);
687 return;
688 }
689 }
690 spin_unlock(&nlm_blocked_lock);
691 printk(KERN_WARNING "lockd: notification for unknown block!\n");
692}
693
694static int nlmsvc_same_owner(struct file_lock *fl1, struct file_lock *fl2)
695{
696 return fl1->fl_owner == fl2->fl_owner && fl1->fl_pid == fl2->fl_pid;
697}
698
699/*
700 * Since NLM uses two "keys" for tracking locks, we need to hash them down
701 * to one for the blocked_hash. Here, we're just xor'ing the host address
702 * with the pid in order to create a key value for picking a hash bucket.
703 */
704static unsigned long
705nlmsvc_owner_key(struct file_lock *fl)
706{
707 return (unsigned long)fl->fl_owner ^ (unsigned long)fl->fl_pid;
708}
709
710const struct lock_manager_operations nlmsvc_lock_operations = {
711 .lm_compare_owner = nlmsvc_same_owner,
712 .lm_owner_key = nlmsvc_owner_key,
713 .lm_notify = nlmsvc_notify_blocked,
714 .lm_grant = nlmsvc_grant_deferred,
715};
716
717/*
718 * Try to claim a lock that was previously blocked.
719 *
720 * Note that we use both the RPC_GRANTED_MSG call _and_ an async
721 * RPC thread when notifying the client. This seems like overkill...
722 * Here's why:
723 * - we don't want to use a synchronous RPC thread, otherwise
724 * we might find ourselves hanging on a dead portmapper.
725 * - Some lockd implementations (e.g. HP) don't react to
726 * RPC_GRANTED calls; they seem to insist on RPC_GRANTED_MSG calls.
727 */
728static void
729nlmsvc_grant_blocked(struct nlm_block *block)
730{
731 struct nlm_file *file = block->b_file;
732 struct nlm_lock *lock = &block->b_call->a_args.lock;
733 int error;
734 loff_t fl_start, fl_end;
735
736 dprintk("lockd: grant blocked lock %p\n", block);
737
738 kref_get(&block->b_count);
739
740 /* Unlink block request from list */
741 nlmsvc_unlink_block(block);
742
743 /* If b_granted is true this means we've been here before.
744 * Just retry the grant callback, possibly refreshing the RPC
745 * binding */
746 if (block->b_granted) {
747 nlm_rebind_host(block->b_host);
748 goto callback;
749 }
750
751 /* Try the lock operation again */
752 /* vfs_lock_file() can mangle fl_start and fl_end, but we need
753 * them unchanged for the GRANT_MSG
754 */
755 lock->fl.fl_flags |= FL_SLEEP;
756 fl_start = lock->fl.fl_start;
757 fl_end = lock->fl.fl_end;
758 error = vfs_lock_file(file->f_file, F_SETLK, &lock->fl, NULL);
759 lock->fl.fl_flags &= ~FL_SLEEP;
760 lock->fl.fl_start = fl_start;
761 lock->fl.fl_end = fl_end;
762
763 switch (error) {
764 case 0:
765 break;
766 case FILE_LOCK_DEFERRED:
767 dprintk("lockd: lock still blocked error %d\n", error);
768 nlmsvc_insert_block(block, NLM_NEVER);
769 nlmsvc_release_block(block);
770 return;
771 default:
772 printk(KERN_WARNING "lockd: unexpected error %d in %s!\n",
773 -error, __func__);
774 nlmsvc_insert_block(block, 10 * HZ);
775 nlmsvc_release_block(block);
776 return;
777 }
778
779callback:
780 /* Lock was granted by VFS. */
781 dprintk("lockd: GRANTing blocked lock.\n");
782 block->b_granted = 1;
783
784 /* keep block on the list, but don't reattempt until the RPC
785 * completes or the submission fails
786 */
787 nlmsvc_insert_block(block, NLM_NEVER);
788
789 /* Call the client -- use a soft RPC task since nlmsvc_retry_blocked
790 * will queue up a new one if this one times out
791 */
792 error = nlm_async_call(block->b_call, NLMPROC_GRANTED_MSG,
793 &nlmsvc_grant_ops);
794
795 /* RPC submission failed, wait a bit and retry */
796 if (error < 0)
797 nlmsvc_insert_block(block, 10 * HZ);
798}
799
800/*
801 * This is the callback from the RPC layer when the NLM_GRANTED_MSG
802 * RPC call has succeeded or timed out.
803 * Like all RPC callbacks, it is invoked by the rpciod process, so it
804 * better not sleep. Therefore, we put the blocked lock on the nlm_blocked
805 * chain once more in order to have it removed by lockd itself (which can
806 * then sleep on the file semaphore without disrupting e.g. the nfs client).
807 */
808static void nlmsvc_grant_callback(struct rpc_task *task, void *data)
809{
810 struct nlm_rqst *call = data;
811 struct nlm_block *block = call->a_block;
812 unsigned long timeout;
813
814 dprintk("lockd: GRANT_MSG RPC callback\n");
815
816 spin_lock(&nlm_blocked_lock);
817 /* if the block is not on a list at this point then it has
818 * been invalidated. Don't try to requeue it.
819 *
820 * FIXME: it's possible that the block is removed from the list
821 * after this check but before the nlmsvc_insert_block. In that
822 * case it will be added back. Perhaps we need better locking
823 * for nlm_blocked?
824 */
825 if (list_empty(&block->b_list))
826 goto out;
827
828 /* Technically, we should down the file semaphore here. Since we
829 * move the block towards the head of the queue only, no harm
830 * can be done, though. */
831 if (task->tk_status < 0) {
832 /* RPC error: Re-insert for retransmission */
833 timeout = 10 * HZ;
834 } else {
835 /* Call was successful, now wait for client callback */
836 timeout = 60 * HZ;
837 }
838 nlmsvc_insert_block_locked(block, timeout);
839 svc_wake_up(block->b_daemon);
840out:
841 spin_unlock(&nlm_blocked_lock);
842}
843
844/*
845 * FIXME: nlmsvc_release_block() grabs a mutex. This is not allowed for an
846 * .rpc_release rpc_call_op
847 */
848static void nlmsvc_grant_release(void *data)
849{
850 struct nlm_rqst *call = data;
851 nlmsvc_release_block(call->a_block);
852}
853
854static const struct rpc_call_ops nlmsvc_grant_ops = {
855 .rpc_call_done = nlmsvc_grant_callback,
856 .rpc_release = nlmsvc_grant_release,
857};
858
859/*
860 * We received a GRANT_RES callback. Try to find the corresponding
861 * block.
862 */
863void
864nlmsvc_grant_reply(struct nlm_cookie *cookie, __be32 status)
865{
866 struct nlm_block *block;
867
868 dprintk("grant_reply: looking for cookie %x, s=%d \n",
869 *(unsigned int *)(cookie->data), status);
870 if (!(block = nlmsvc_find_block(cookie)))
871 return;
872
873 if (block) {
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 /* Lock is now held by client, or has been rejected.
879 * In both cases, the block should be removed. */
880 nlmsvc_unlink_block(block);
881 }
882 }
883 nlmsvc_release_block(block);
884}
885
886/* Helper function to handle retry of a deferred block.
887 * If it is a blocking lock, call grant_blocked.
888 * For a non-blocking lock or test lock, revisit the request.
889 */
890static void
891retry_deferred_block(struct nlm_block *block)
892{
893 if (!(block->b_flags & B_GOT_CALLBACK))
894 block->b_flags |= B_TIMED_OUT;
895 nlmsvc_insert_block(block, NLM_TIMEOUT);
896 dprintk("revisit block %p flags %d\n", block, block->b_flags);
897 if (block->b_deferred_req) {
898 block->b_deferred_req->revisit(block->b_deferred_req, 0);
899 block->b_deferred_req = NULL;
900 }
901}
902
903/*
904 * Retry all blocked locks that have been notified. This is where lockd
905 * picks up locks that can be granted, or grant notifications that must
906 * be retransmitted.
907 */
908unsigned long
909nlmsvc_retry_blocked(void)
910{
911 unsigned long timeout = MAX_SCHEDULE_TIMEOUT;
912 struct nlm_block *block;
913
914 spin_lock(&nlm_blocked_lock);
915 while (!list_empty(&nlm_blocked) && !kthread_should_stop()) {
916 block = list_entry(nlm_blocked.next, struct nlm_block, b_list);
917
918 if (block->b_when == NLM_NEVER)
919 break;
920 if (time_after(block->b_when, jiffies)) {
921 timeout = block->b_when - jiffies;
922 break;
923 }
924 spin_unlock(&nlm_blocked_lock);
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 spin_lock(&nlm_blocked_lock);
935 }
936 spin_unlock(&nlm_blocked_lock);
937
938 return timeout;
939}
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/exportfs.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 spin_lock(&nlm_blocked_lock);
134 if (!list_empty(&block->b_list)) {
135 list_del_init(&block->b_list);
136 spin_unlock(&nlm_blocked_lock);
137 nlmsvc_release_block(block);
138 return;
139 }
140 spin_unlock(&nlm_blocked_lock);
141}
142
143/*
144 * Find a block for a given lock
145 */
146static struct nlm_block *
147nlmsvc_lookup_block(struct nlm_file *file, struct nlm_lock *lock)
148{
149 struct nlm_block *block;
150 struct file_lock *fl;
151
152 dprintk("lockd: nlmsvc_lookup_block f=%p pd=%d %Ld-%Ld ty=%d\n",
153 file, lock->fl.fl_pid,
154 (long long)lock->fl.fl_start,
155 (long long)lock->fl.fl_end, lock->fl.fl_type);
156 spin_lock(&nlm_blocked_lock);
157 list_for_each_entry(block, &nlm_blocked, b_list) {
158 fl = &block->b_call->a_args.lock.fl;
159 dprintk("lockd: check f=%p pd=%d %Ld-%Ld ty=%d cookie=%s\n",
160 block->b_file, fl->fl_pid,
161 (long long)fl->fl_start,
162 (long long)fl->fl_end, fl->fl_type,
163 nlmdbg_cookie2a(&block->b_call->a_args.cookie));
164 if (block->b_file == file && nlm_compare_locks(fl, &lock->fl)) {
165 kref_get(&block->b_count);
166 spin_unlock(&nlm_blocked_lock);
167 return block;
168 }
169 }
170 spin_unlock(&nlm_blocked_lock);
171
172 return NULL;
173}
174
175static inline int nlm_cookie_match(struct nlm_cookie *a, struct nlm_cookie *b)
176{
177 if (a->len != b->len)
178 return 0;
179 if (memcmp(a->data, b->data, a->len))
180 return 0;
181 return 1;
182}
183
184/*
185 * Find a block with a given NLM cookie.
186 */
187static inline struct nlm_block *
188nlmsvc_find_block(struct nlm_cookie *cookie)
189{
190 struct nlm_block *block;
191
192 spin_lock(&nlm_blocked_lock);
193 list_for_each_entry(block, &nlm_blocked, b_list) {
194 if (nlm_cookie_match(&block->b_call->a_args.cookie,cookie))
195 goto found;
196 }
197 spin_unlock(&nlm_blocked_lock);
198
199 return NULL;
200
201found:
202 dprintk("nlmsvc_find_block(%s): block=%p\n", nlmdbg_cookie2a(cookie), block);
203 kref_get(&block->b_count);
204 spin_unlock(&nlm_blocked_lock);
205 return block;
206}
207
208/*
209 * Create a block and initialize it.
210 *
211 * Note: we explicitly set the cookie of the grant reply to that of
212 * the blocked lock request. The spec explicitly mentions that the client
213 * should _not_ rely on the callback containing the same cookie as the
214 * request, but (as I found out later) that's because some implementations
215 * do just this. Never mind the standards comittees, they support our
216 * logging industries.
217 *
218 * 10 years later: I hope we can safely ignore these old and broken
219 * clients by now. Let's fix this so we can uniquely identify an incoming
220 * GRANTED_RES message by cookie, without having to rely on the client's IP
221 * address. --okir
222 */
223static struct nlm_block *
224nlmsvc_create_block(struct svc_rqst *rqstp, struct nlm_host *host,
225 struct nlm_file *file, struct nlm_lock *lock,
226 struct nlm_cookie *cookie)
227{
228 struct nlm_block *block;
229 struct nlm_rqst *call = NULL;
230
231 call = nlm_alloc_call(host);
232 if (call == NULL)
233 return NULL;
234
235 /* Allocate memory for block, and initialize arguments */
236 block = kzalloc(sizeof(*block), GFP_KERNEL);
237 if (block == NULL)
238 goto failed;
239 kref_init(&block->b_count);
240 INIT_LIST_HEAD(&block->b_list);
241 INIT_LIST_HEAD(&block->b_flist);
242
243 if (!nlmsvc_setgrantargs(call, lock))
244 goto failed_free;
245
246 /* Set notifier function for VFS, and init args */
247 call->a_args.lock.fl.fl_flags |= FL_SLEEP;
248 call->a_args.lock.fl.fl_lmops = &nlmsvc_lock_operations;
249 nlmclnt_next_cookie(&call->a_args.cookie);
250
251 dprintk("lockd: created block %p...\n", block);
252
253 /* Create and initialize the block */
254 block->b_daemon = rqstp->rq_server;
255 block->b_host = host;
256 block->b_file = file;
257 file->f_count++;
258
259 /* Add to file's list of blocks */
260 list_add(&block->b_flist, &file->f_blocks);
261
262 /* Set up RPC arguments for callback */
263 block->b_call = call;
264 call->a_flags = RPC_TASK_ASYNC;
265 call->a_block = block;
266
267 return block;
268
269failed_free:
270 kfree(block);
271failed:
272 nlmsvc_release_call(call);
273 return NULL;
274}
275
276/*
277 * Delete a block.
278 * It is the caller's responsibility to check whether the file
279 * can be closed hereafter.
280 */
281static int nlmsvc_unlink_block(struct nlm_block *block)
282{
283 int status;
284 dprintk("lockd: unlinking block %p...\n", block);
285
286 /* Remove block from list */
287 status = locks_delete_block(&block->b_call->a_args.lock.fl);
288 nlmsvc_remove_block(block);
289 return status;
290}
291
292static void nlmsvc_free_block(struct kref *kref)
293{
294 struct nlm_block *block = container_of(kref, struct nlm_block, b_count);
295 struct nlm_file *file = block->b_file;
296
297 dprintk("lockd: freeing block %p...\n", block);
298
299 /* Remove block from file's list of blocks */
300 list_del_init(&block->b_flist);
301 mutex_unlock(&file->f_mutex);
302
303 nlmsvc_freegrantargs(block->b_call);
304 nlmsvc_release_call(block->b_call);
305 nlm_release_file(block->b_file);
306 kfree(block);
307}
308
309static void nlmsvc_release_block(struct nlm_block *block)
310{
311 if (block != NULL)
312 kref_put_mutex(&block->b_count, nlmsvc_free_block, &block->b_file->f_mutex);
313}
314
315/*
316 * Loop over all blocks and delete blocks held by
317 * a matching host.
318 */
319void nlmsvc_traverse_blocks(struct nlm_host *host,
320 struct nlm_file *file,
321 nlm_host_match_fn_t match)
322{
323 struct nlm_block *block, *next;
324
325restart:
326 mutex_lock(&file->f_mutex);
327 spin_lock(&nlm_blocked_lock);
328 list_for_each_entry_safe(block, next, &file->f_blocks, b_flist) {
329 if (!match(block->b_host, host))
330 continue;
331 /* Do not destroy blocks that are not on
332 * the global retry list - why? */
333 if (list_empty(&block->b_list))
334 continue;
335 kref_get(&block->b_count);
336 spin_unlock(&nlm_blocked_lock);
337 mutex_unlock(&file->f_mutex);
338 nlmsvc_unlink_block(block);
339 nlmsvc_release_block(block);
340 goto restart;
341 }
342 spin_unlock(&nlm_blocked_lock);
343 mutex_unlock(&file->f_mutex);
344}
345
346static struct nlm_lockowner *
347nlmsvc_get_lockowner(struct nlm_lockowner *lockowner)
348{
349 refcount_inc(&lockowner->count);
350 return lockowner;
351}
352
353void nlmsvc_put_lockowner(struct nlm_lockowner *lockowner)
354{
355 if (!refcount_dec_and_lock(&lockowner->count, &lockowner->host->h_lock))
356 return;
357 list_del(&lockowner->list);
358 spin_unlock(&lockowner->host->h_lock);
359 nlmsvc_release_host(lockowner->host);
360 kfree(lockowner);
361}
362
363static struct nlm_lockowner *__nlmsvc_find_lockowner(struct nlm_host *host, pid_t pid)
364{
365 struct nlm_lockowner *lockowner;
366 list_for_each_entry(lockowner, &host->h_lockowners, list) {
367 if (lockowner->pid != pid)
368 continue;
369 return nlmsvc_get_lockowner(lockowner);
370 }
371 return NULL;
372}
373
374static struct nlm_lockowner *nlmsvc_find_lockowner(struct nlm_host *host, pid_t pid)
375{
376 struct nlm_lockowner *res, *new = NULL;
377
378 spin_lock(&host->h_lock);
379 res = __nlmsvc_find_lockowner(host, pid);
380
381 if (res == NULL) {
382 spin_unlock(&host->h_lock);
383 new = kmalloc(sizeof(*res), GFP_KERNEL);
384 spin_lock(&host->h_lock);
385 res = __nlmsvc_find_lockowner(host, pid);
386 if (res == NULL && new != NULL) {
387 res = new;
388 /* fs/locks.c will manage the refcount through lock_ops */
389 refcount_set(&new->count, 1);
390 new->pid = pid;
391 new->host = nlm_get_host(host);
392 list_add(&new->list, &host->h_lockowners);
393 new = NULL;
394 }
395 }
396
397 spin_unlock(&host->h_lock);
398 kfree(new);
399 return res;
400}
401
402void
403nlmsvc_release_lockowner(struct nlm_lock *lock)
404{
405 if (lock->fl.fl_owner)
406 nlmsvc_put_lockowner(lock->fl.fl_owner);
407}
408
409void nlmsvc_locks_init_private(struct file_lock *fl, struct nlm_host *host,
410 pid_t pid)
411{
412 fl->fl_owner = nlmsvc_find_lockowner(host, pid);
413}
414
415/*
416 * Initialize arguments for GRANTED call. The nlm_rqst structure
417 * has been cleared already.
418 */
419static int nlmsvc_setgrantargs(struct nlm_rqst *call, struct nlm_lock *lock)
420{
421 locks_copy_lock(&call->a_args.lock.fl, &lock->fl);
422 memcpy(&call->a_args.lock.fh, &lock->fh, sizeof(call->a_args.lock.fh));
423 call->a_args.lock.caller = utsname()->nodename;
424 call->a_args.lock.oh.len = lock->oh.len;
425
426 /* set default data area */
427 call->a_args.lock.oh.data = call->a_owner;
428 call->a_args.lock.svid = ((struct nlm_lockowner *)lock->fl.fl_owner)->pid;
429
430 if (lock->oh.len > NLMCLNT_OHSIZE) {
431 void *data = kmalloc(lock->oh.len, GFP_KERNEL);
432 if (!data)
433 return 0;
434 call->a_args.lock.oh.data = (u8 *) data;
435 }
436
437 memcpy(call->a_args.lock.oh.data, lock->oh.data, lock->oh.len);
438 return 1;
439}
440
441static void nlmsvc_freegrantargs(struct nlm_rqst *call)
442{
443 if (call->a_args.lock.oh.data != call->a_owner)
444 kfree(call->a_args.lock.oh.data);
445
446 locks_release_private(&call->a_args.lock.fl);
447}
448
449/*
450 * Deferred lock request handling for non-blocking lock
451 */
452static __be32
453nlmsvc_defer_lock_rqst(struct svc_rqst *rqstp, struct nlm_block *block)
454{
455 __be32 status = nlm_lck_denied_nolocks;
456
457 block->b_flags |= B_QUEUED;
458
459 nlmsvc_insert_block(block, NLM_TIMEOUT);
460
461 block->b_cache_req = &rqstp->rq_chandle;
462 if (rqstp->rq_chandle.defer) {
463 block->b_deferred_req =
464 rqstp->rq_chandle.defer(block->b_cache_req);
465 if (block->b_deferred_req != NULL)
466 status = nlm_drop_reply;
467 }
468 dprintk("lockd: nlmsvc_defer_lock_rqst block %p flags %d status %d\n",
469 block, block->b_flags, ntohl(status));
470
471 return status;
472}
473
474/*
475 * Attempt to establish a lock, and if it can't be granted, block it
476 * if required.
477 */
478__be32
479nlmsvc_lock(struct svc_rqst *rqstp, struct nlm_file *file,
480 struct nlm_host *host, struct nlm_lock *lock, int wait,
481 struct nlm_cookie *cookie, int reclaim)
482{
483 struct inode *inode = nlmsvc_file_inode(file);
484 struct nlm_block *block = NULL;
485 int error;
486 int mode;
487 int async_block = 0;
488 __be32 ret;
489
490 dprintk("lockd: nlmsvc_lock(%s/%ld, ty=%d, pi=%d, %Ld-%Ld, bl=%d)\n",
491 inode->i_sb->s_id, inode->i_ino,
492 lock->fl.fl_type, lock->fl.fl_pid,
493 (long long)lock->fl.fl_start,
494 (long long)lock->fl.fl_end,
495 wait);
496
497 if (!exportfs_lock_op_is_async(inode->i_sb->s_export_op)) {
498 async_block = wait;
499 wait = 0;
500 }
501
502 /* Lock file against concurrent access */
503 mutex_lock(&file->f_mutex);
504 /* Get existing block (in case client is busy-waiting)
505 * or create new block
506 */
507 block = nlmsvc_lookup_block(file, lock);
508 if (block == NULL) {
509 block = nlmsvc_create_block(rqstp, host, file, lock, cookie);
510 ret = nlm_lck_denied_nolocks;
511 if (block == NULL)
512 goto out;
513 lock = &block->b_call->a_args.lock;
514 } else
515 lock->fl.fl_flags &= ~FL_SLEEP;
516
517 if (block->b_flags & B_QUEUED) {
518 dprintk("lockd: nlmsvc_lock deferred block %p flags %d\n",
519 block, block->b_flags);
520 if (block->b_granted) {
521 nlmsvc_unlink_block(block);
522 ret = nlm_granted;
523 goto out;
524 }
525 if (block->b_flags & B_TIMED_OUT) {
526 nlmsvc_unlink_block(block);
527 ret = nlm_lck_denied;
528 goto out;
529 }
530 ret = nlm_drop_reply;
531 goto out;
532 }
533
534 if (locks_in_grace(SVC_NET(rqstp)) && !reclaim) {
535 ret = nlm_lck_denied_grace_period;
536 goto out;
537 }
538 if (reclaim && !locks_in_grace(SVC_NET(rqstp))) {
539 ret = nlm_lck_denied_grace_period;
540 goto out;
541 }
542
543 spin_lock(&nlm_blocked_lock);
544 /*
545 * If this is a lock request for an already pending
546 * lock request we return nlm_lck_blocked without calling
547 * vfs_lock_file() again. Otherwise we have two pending
548 * requests on the underlaying ->lock() implementation but
549 * only one nlm_block to being granted by lm_grant().
550 */
551 if (exportfs_lock_op_is_async(inode->i_sb->s_export_op) &&
552 !list_empty(&block->b_list)) {
553 spin_unlock(&nlm_blocked_lock);
554 ret = nlm_lck_blocked;
555 goto out;
556 }
557
558 /* Append to list of blocked */
559 nlmsvc_insert_block_locked(block, NLM_NEVER);
560 spin_unlock(&nlm_blocked_lock);
561
562 if (!wait)
563 lock->fl.fl_flags &= ~FL_SLEEP;
564 mode = lock_to_openmode(&lock->fl);
565 error = vfs_lock_file(file->f_file[mode], F_SETLK, &lock->fl, NULL);
566 lock->fl.fl_flags &= ~FL_SLEEP;
567
568 dprintk("lockd: vfs_lock_file returned %d\n", error);
569 switch (error) {
570 case 0:
571 nlmsvc_remove_block(block);
572 ret = nlm_granted;
573 goto out;
574 case -EAGAIN:
575 if (!wait)
576 nlmsvc_remove_block(block);
577 ret = async_block ? nlm_lck_blocked : nlm_lck_denied;
578 goto out;
579 case FILE_LOCK_DEFERRED:
580 if (wait)
581 break;
582 /* Filesystem lock operation is in progress
583 Add it to the queue waiting for callback */
584 ret = nlmsvc_defer_lock_rqst(rqstp, block);
585 goto out;
586 case -EDEADLK:
587 nlmsvc_remove_block(block);
588 ret = nlm_deadlock;
589 goto out;
590 default: /* includes ENOLCK */
591 nlmsvc_remove_block(block);
592 ret = nlm_lck_denied_nolocks;
593 goto out;
594 }
595
596 ret = nlm_lck_blocked;
597out:
598 mutex_unlock(&file->f_mutex);
599 nlmsvc_release_block(block);
600 dprintk("lockd: nlmsvc_lock returned %u\n", ret);
601 return ret;
602}
603
604/*
605 * Test for presence of a conflicting lock.
606 */
607__be32
608nlmsvc_testlock(struct svc_rqst *rqstp, struct nlm_file *file,
609 struct nlm_host *host, struct nlm_lock *lock,
610 struct nlm_lock *conflock, struct nlm_cookie *cookie)
611{
612 int error;
613 int mode;
614 __be32 ret;
615
616 dprintk("lockd: nlmsvc_testlock(%s/%ld, ty=%d, %Ld-%Ld)\n",
617 nlmsvc_file_inode(file)->i_sb->s_id,
618 nlmsvc_file_inode(file)->i_ino,
619 lock->fl.fl_type,
620 (long long)lock->fl.fl_start,
621 (long long)lock->fl.fl_end);
622
623 if (locks_in_grace(SVC_NET(rqstp))) {
624 ret = nlm_lck_denied_grace_period;
625 goto out;
626 }
627
628 mode = lock_to_openmode(&lock->fl);
629 error = vfs_test_lock(file->f_file[mode], &lock->fl);
630 if (error) {
631 /* We can't currently deal with deferred test requests */
632 if (error == FILE_LOCK_DEFERRED)
633 WARN_ON_ONCE(1);
634
635 ret = nlm_lck_denied_nolocks;
636 goto out;
637 }
638
639 if (lock->fl.fl_type == F_UNLCK) {
640 ret = nlm_granted;
641 goto out;
642 }
643
644 dprintk("lockd: conflicting lock(ty=%d, %Ld-%Ld)\n",
645 lock->fl.fl_type, (long long)lock->fl.fl_start,
646 (long long)lock->fl.fl_end);
647 conflock->caller = "somehost"; /* FIXME */
648 conflock->len = strlen(conflock->caller);
649 conflock->oh.len = 0; /* don't return OH info */
650 conflock->svid = lock->fl.fl_pid;
651 conflock->fl.fl_type = lock->fl.fl_type;
652 conflock->fl.fl_start = lock->fl.fl_start;
653 conflock->fl.fl_end = lock->fl.fl_end;
654 locks_release_private(&lock->fl);
655
656 ret = nlm_lck_denied;
657out:
658 return ret;
659}
660
661/*
662 * Remove a lock.
663 * This implies a CANCEL call: We send a GRANT_MSG, the client replies
664 * with a GRANT_RES call which gets lost, and calls UNLOCK immediately
665 * afterwards. In this case the block will still be there, and hence
666 * must be removed.
667 */
668__be32
669nlmsvc_unlock(struct net *net, struct nlm_file *file, struct nlm_lock *lock)
670{
671 int error = 0;
672
673 dprintk("lockd: nlmsvc_unlock(%s/%ld, pi=%d, %Ld-%Ld)\n",
674 nlmsvc_file_inode(file)->i_sb->s_id,
675 nlmsvc_file_inode(file)->i_ino,
676 lock->fl.fl_pid,
677 (long long)lock->fl.fl_start,
678 (long long)lock->fl.fl_end);
679
680 /* First, cancel any lock that might be there */
681 nlmsvc_cancel_blocked(net, file, lock);
682
683 lock->fl.fl_type = F_UNLCK;
684 lock->fl.fl_file = file->f_file[O_RDONLY];
685 if (lock->fl.fl_file)
686 error = vfs_lock_file(lock->fl.fl_file, F_SETLK,
687 &lock->fl, NULL);
688 lock->fl.fl_file = file->f_file[O_WRONLY];
689 if (lock->fl.fl_file)
690 error |= vfs_lock_file(lock->fl.fl_file, F_SETLK,
691 &lock->fl, NULL);
692
693 return (error < 0)? nlm_lck_denied_nolocks : nlm_granted;
694}
695
696/*
697 * Cancel a previously blocked request.
698 *
699 * A cancel request always overrides any grant that may currently
700 * be in progress.
701 * The calling procedure must check whether the file can be closed.
702 */
703__be32
704nlmsvc_cancel_blocked(struct net *net, struct nlm_file *file, struct nlm_lock *lock)
705{
706 struct nlm_block *block;
707 int status = 0;
708 int mode;
709
710 dprintk("lockd: nlmsvc_cancel(%s/%ld, pi=%d, %Ld-%Ld)\n",
711 nlmsvc_file_inode(file)->i_sb->s_id,
712 nlmsvc_file_inode(file)->i_ino,
713 lock->fl.fl_pid,
714 (long long)lock->fl.fl_start,
715 (long long)lock->fl.fl_end);
716
717 if (locks_in_grace(net))
718 return nlm_lck_denied_grace_period;
719
720 mutex_lock(&file->f_mutex);
721 block = nlmsvc_lookup_block(file, lock);
722 mutex_unlock(&file->f_mutex);
723 if (block != NULL) {
724 struct file_lock *fl = &block->b_call->a_args.lock.fl;
725
726 mode = lock_to_openmode(fl);
727 vfs_cancel_lock(block->b_file->f_file[mode], fl);
728 status = nlmsvc_unlink_block(block);
729 nlmsvc_release_block(block);
730 }
731 return status ? nlm_lck_denied : nlm_granted;
732}
733
734/*
735 * This is a callback from the filesystem for VFS file lock requests.
736 * It will be used if lm_grant is defined and the filesystem can not
737 * respond to the request immediately.
738 * For SETLK or SETLKW request it will get the local posix lock.
739 * In all cases it will move the block to the head of nlm_blocked q where
740 * nlmsvc_retry_blocked() can send back a reply for SETLKW or revisit the
741 * deferred rpc for GETLK and SETLK.
742 */
743static void
744nlmsvc_update_deferred_block(struct nlm_block *block, int result)
745{
746 block->b_flags |= B_GOT_CALLBACK;
747 if (result == 0)
748 block->b_granted = 1;
749 else
750 block->b_flags |= B_TIMED_OUT;
751}
752
753static int nlmsvc_grant_deferred(struct file_lock *fl, int result)
754{
755 struct nlm_block *block;
756 int rc = -ENOENT;
757
758 spin_lock(&nlm_blocked_lock);
759 list_for_each_entry(block, &nlm_blocked, b_list) {
760 if (nlm_compare_locks(&block->b_call->a_args.lock.fl, fl)) {
761 dprintk("lockd: nlmsvc_notify_blocked block %p flags %d\n",
762 block, block->b_flags);
763 if (block->b_flags & B_QUEUED) {
764 if (block->b_flags & B_TIMED_OUT) {
765 rc = -ENOLCK;
766 break;
767 }
768 nlmsvc_update_deferred_block(block, result);
769 } else if (result == 0)
770 block->b_granted = 1;
771
772 nlmsvc_insert_block_locked(block, 0);
773 svc_wake_up(block->b_daemon);
774 rc = 0;
775 break;
776 }
777 }
778 spin_unlock(&nlm_blocked_lock);
779 if (rc == -ENOENT)
780 printk(KERN_WARNING "lockd: grant for unknown block\n");
781 return rc;
782}
783
784/*
785 * Unblock a blocked lock request. This is a callback invoked from the
786 * VFS layer when a lock on which we blocked is removed.
787 *
788 * This function doesn't grant the blocked lock instantly, but rather moves
789 * the block to the head of nlm_blocked where it can be picked up by lockd.
790 */
791static void
792nlmsvc_notify_blocked(struct file_lock *fl)
793{
794 struct nlm_block *block;
795
796 dprintk("lockd: VFS unblock notification for block %p\n", fl);
797 spin_lock(&nlm_blocked_lock);
798 list_for_each_entry(block, &nlm_blocked, b_list) {
799 if (nlm_compare_locks(&block->b_call->a_args.lock.fl, fl)) {
800 nlmsvc_insert_block_locked(block, 0);
801 spin_unlock(&nlm_blocked_lock);
802 svc_wake_up(block->b_daemon);
803 return;
804 }
805 }
806 spin_unlock(&nlm_blocked_lock);
807 printk(KERN_WARNING "lockd: notification for unknown block!\n");
808}
809
810static fl_owner_t nlmsvc_get_owner(fl_owner_t owner)
811{
812 return nlmsvc_get_lockowner(owner);
813}
814
815static void nlmsvc_put_owner(fl_owner_t owner)
816{
817 nlmsvc_put_lockowner(owner);
818}
819
820const struct lock_manager_operations nlmsvc_lock_operations = {
821 .lm_notify = nlmsvc_notify_blocked,
822 .lm_grant = nlmsvc_grant_deferred,
823 .lm_get_owner = nlmsvc_get_owner,
824 .lm_put_owner = nlmsvc_put_owner,
825};
826
827/*
828 * Try to claim a lock that was previously blocked.
829 *
830 * Note that we use both the RPC_GRANTED_MSG call _and_ an async
831 * RPC thread when notifying the client. This seems like overkill...
832 * Here's why:
833 * - we don't want to use a synchronous RPC thread, otherwise
834 * we might find ourselves hanging on a dead portmapper.
835 * - Some lockd implementations (e.g. HP) don't react to
836 * RPC_GRANTED calls; they seem to insist on RPC_GRANTED_MSG calls.
837 */
838static void
839nlmsvc_grant_blocked(struct nlm_block *block)
840{
841 struct nlm_file *file = block->b_file;
842 struct nlm_lock *lock = &block->b_call->a_args.lock;
843 int mode;
844 int error;
845 loff_t fl_start, fl_end;
846
847 dprintk("lockd: grant blocked lock %p\n", block);
848
849 kref_get(&block->b_count);
850
851 /* Unlink block request from list */
852 nlmsvc_unlink_block(block);
853
854 /* If b_granted is true this means we've been here before.
855 * Just retry the grant callback, possibly refreshing the RPC
856 * binding */
857 if (block->b_granted) {
858 nlm_rebind_host(block->b_host);
859 goto callback;
860 }
861
862 /* Try the lock operation again */
863 /* vfs_lock_file() can mangle fl_start and fl_end, but we need
864 * them unchanged for the GRANT_MSG
865 */
866 lock->fl.fl_flags |= FL_SLEEP;
867 fl_start = lock->fl.fl_start;
868 fl_end = lock->fl.fl_end;
869 mode = lock_to_openmode(&lock->fl);
870 error = vfs_lock_file(file->f_file[mode], F_SETLK, &lock->fl, NULL);
871 lock->fl.fl_flags &= ~FL_SLEEP;
872 lock->fl.fl_start = fl_start;
873 lock->fl.fl_end = fl_end;
874
875 switch (error) {
876 case 0:
877 break;
878 case FILE_LOCK_DEFERRED:
879 dprintk("lockd: lock still blocked error %d\n", error);
880 nlmsvc_insert_block(block, NLM_NEVER);
881 nlmsvc_release_block(block);
882 return;
883 default:
884 printk(KERN_WARNING "lockd: unexpected error %d in %s!\n",
885 -error, __func__);
886 nlmsvc_insert_block(block, 10 * HZ);
887 nlmsvc_release_block(block);
888 return;
889 }
890
891callback:
892 /* Lock was granted by VFS. */
893 dprintk("lockd: GRANTing blocked lock.\n");
894 block->b_granted = 1;
895
896 /* keep block on the list, but don't reattempt until the RPC
897 * completes or the submission fails
898 */
899 nlmsvc_insert_block(block, NLM_NEVER);
900
901 /* Call the client -- use a soft RPC task since nlmsvc_retry_blocked
902 * will queue up a new one if this one times out
903 */
904 error = nlm_async_call(block->b_call, NLMPROC_GRANTED_MSG,
905 &nlmsvc_grant_ops);
906
907 /* RPC submission failed, wait a bit and retry */
908 if (error < 0)
909 nlmsvc_insert_block(block, 10 * HZ);
910}
911
912/*
913 * This is the callback from the RPC layer when the NLM_GRANTED_MSG
914 * RPC call has succeeded or timed out.
915 * Like all RPC callbacks, it is invoked by the rpciod process, so it
916 * better not sleep. Therefore, we put the blocked lock on the nlm_blocked
917 * chain once more in order to have it removed by lockd itself (which can
918 * then sleep on the file semaphore without disrupting e.g. the nfs client).
919 */
920static void nlmsvc_grant_callback(struct rpc_task *task, void *data)
921{
922 struct nlm_rqst *call = data;
923 struct nlm_block *block = call->a_block;
924 unsigned long timeout;
925
926 dprintk("lockd: GRANT_MSG RPC callback\n");
927
928 spin_lock(&nlm_blocked_lock);
929 /* if the block is not on a list at this point then it has
930 * been invalidated. Don't try to requeue it.
931 *
932 * FIXME: it's possible that the block is removed from the list
933 * after this check but before the nlmsvc_insert_block. In that
934 * case it will be added back. Perhaps we need better locking
935 * for nlm_blocked?
936 */
937 if (list_empty(&block->b_list))
938 goto out;
939
940 /* Technically, we should down the file semaphore here. Since we
941 * move the block towards the head of the queue only, no harm
942 * can be done, though. */
943 if (task->tk_status < 0) {
944 /* RPC error: Re-insert for retransmission */
945 timeout = 10 * HZ;
946 } else {
947 /* Call was successful, now wait for client callback */
948 timeout = 60 * HZ;
949 }
950 nlmsvc_insert_block_locked(block, timeout);
951 svc_wake_up(block->b_daemon);
952out:
953 spin_unlock(&nlm_blocked_lock);
954}
955
956/*
957 * FIXME: nlmsvc_release_block() grabs a mutex. This is not allowed for an
958 * .rpc_release rpc_call_op
959 */
960static void nlmsvc_grant_release(void *data)
961{
962 struct nlm_rqst *call = data;
963 nlmsvc_release_block(call->a_block);
964}
965
966static const struct rpc_call_ops nlmsvc_grant_ops = {
967 .rpc_call_done = nlmsvc_grant_callback,
968 .rpc_release = nlmsvc_grant_release,
969};
970
971/*
972 * We received a GRANT_RES callback. Try to find the corresponding
973 * block.
974 */
975void
976nlmsvc_grant_reply(struct nlm_cookie *cookie, __be32 status)
977{
978 struct nlm_block *block;
979 struct file_lock *fl;
980 int error;
981
982 dprintk("grant_reply: looking for cookie %x, s=%d \n",
983 *(unsigned int *)(cookie->data), status);
984 if (!(block = nlmsvc_find_block(cookie)))
985 return;
986
987 switch (status) {
988 case nlm_lck_denied_grace_period:
989 /* Try again in a couple of seconds */
990 nlmsvc_insert_block(block, 10 * HZ);
991 break;
992 case nlm_lck_denied:
993 /* Client doesn't want it, just unlock it */
994 nlmsvc_unlink_block(block);
995 fl = &block->b_call->a_args.lock.fl;
996 fl->fl_type = F_UNLCK;
997 error = vfs_lock_file(fl->fl_file, F_SETLK, fl, NULL);
998 if (error)
999 pr_warn("lockd: unable to unlock lock rejected by client!\n");
1000 break;
1001 default:
1002 /*
1003 * Either it was accepted or the status makes no sense
1004 * just unlink it either way.
1005 */
1006 nlmsvc_unlink_block(block);
1007 }
1008 nlmsvc_release_block(block);
1009}
1010
1011/* Helper function to handle retry of a deferred block.
1012 * If it is a blocking lock, call grant_blocked.
1013 * For a non-blocking lock or test lock, revisit the request.
1014 */
1015static void
1016retry_deferred_block(struct nlm_block *block)
1017{
1018 if (!(block->b_flags & B_GOT_CALLBACK))
1019 block->b_flags |= B_TIMED_OUT;
1020 nlmsvc_insert_block(block, NLM_TIMEOUT);
1021 dprintk("revisit block %p flags %d\n", block, block->b_flags);
1022 if (block->b_deferred_req) {
1023 block->b_deferred_req->revisit(block->b_deferred_req, 0);
1024 block->b_deferred_req = NULL;
1025 }
1026}
1027
1028/*
1029 * Retry all blocked locks that have been notified. This is where lockd
1030 * picks up locks that can be granted, or grant notifications that must
1031 * be retransmitted.
1032 */
1033void
1034nlmsvc_retry_blocked(struct svc_rqst *rqstp)
1035{
1036 unsigned long timeout = MAX_SCHEDULE_TIMEOUT;
1037 struct nlm_block *block;
1038
1039 spin_lock(&nlm_blocked_lock);
1040 while (!list_empty(&nlm_blocked) && !svc_thread_should_stop(rqstp)) {
1041 block = list_entry(nlm_blocked.next, struct nlm_block, b_list);
1042
1043 if (block->b_when == NLM_NEVER)
1044 break;
1045 if (time_after(block->b_when, jiffies)) {
1046 timeout = block->b_when - jiffies;
1047 break;
1048 }
1049 spin_unlock(&nlm_blocked_lock);
1050
1051 dprintk("nlmsvc_retry_blocked(%p, when=%ld)\n",
1052 block, block->b_when);
1053 if (block->b_flags & B_QUEUED) {
1054 dprintk("nlmsvc_retry_blocked delete block (%p, granted=%d, flags=%d)\n",
1055 block, block->b_granted, block->b_flags);
1056 retry_deferred_block(block);
1057 } else
1058 nlmsvc_grant_blocked(block);
1059 spin_lock(&nlm_blocked_lock);
1060 }
1061 spin_unlock(&nlm_blocked_lock);
1062
1063 if (timeout < MAX_SCHEDULE_TIMEOUT)
1064 mod_timer(&nlmsvc_retry, jiffies + timeout);
1065}