Loading...
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#include <linux/exportfs.h>
35
36#define NLMDBG_FACILITY NLMDBG_SVCLOCK
37
38#ifdef CONFIG_LOCKD_V4
39#define nlm_deadlock nlm4_deadlock
40#else
41#define nlm_deadlock nlm_lck_denied
42#endif
43
44static void nlmsvc_release_block(struct nlm_block *block);
45static void nlmsvc_insert_block(struct nlm_block *block, unsigned long);
46static void nlmsvc_remove_block(struct nlm_block *block);
47
48static int nlmsvc_setgrantargs(struct nlm_rqst *call, struct nlm_lock *lock);
49static void nlmsvc_freegrantargs(struct nlm_rqst *call);
50static const struct rpc_call_ops nlmsvc_grant_ops;
51
52/*
53 * The list of blocked locks to retry
54 */
55static LIST_HEAD(nlm_blocked);
56static DEFINE_SPINLOCK(nlm_blocked_lock);
57
58#if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
59static const char *nlmdbg_cookie2a(const struct nlm_cookie *cookie)
60{
61 /*
62 * We can get away with a static buffer because this is only called
63 * from lockd, which is single-threaded.
64 */
65 static char buf[2*NLM_MAXCOOKIELEN+1];
66 unsigned int i, len = sizeof(buf);
67 char *p = buf;
68
69 len--; /* allow for trailing \0 */
70 if (len < 3)
71 return "???";
72 for (i = 0 ; i < cookie->len ; i++) {
73 if (len < 2) {
74 strcpy(p-3, "...");
75 break;
76 }
77 sprintf(p, "%02x", cookie->data[i]);
78 p += 2;
79 len -= 2;
80 }
81 *p = '\0';
82
83 return buf;
84}
85#endif
86
87/*
88 * Insert a blocked lock into the global list
89 */
90static void
91nlmsvc_insert_block_locked(struct nlm_block *block, unsigned long when)
92{
93 struct nlm_block *b;
94 struct list_head *pos;
95
96 dprintk("lockd: nlmsvc_insert_block(%p, %ld)\n", block, when);
97 if (list_empty(&block->b_list)) {
98 kref_get(&block->b_count);
99 } else {
100 list_del_init(&block->b_list);
101 }
102
103 pos = &nlm_blocked;
104 if (when != NLM_NEVER) {
105 if ((when += jiffies) == NLM_NEVER)
106 when ++;
107 list_for_each(pos, &nlm_blocked) {
108 b = list_entry(pos, struct nlm_block, b_list);
109 if (time_after(b->b_when,when) || b->b_when == NLM_NEVER)
110 break;
111 }
112 /* On normal exit from the loop, pos == &nlm_blocked,
113 * so we will be adding to the end of the list - good
114 */
115 }
116
117 list_add_tail(&block->b_list, pos);
118 block->b_when = when;
119}
120
121static void nlmsvc_insert_block(struct nlm_block *block, unsigned long when)
122{
123 spin_lock(&nlm_blocked_lock);
124 nlmsvc_insert_block_locked(block, when);
125 spin_unlock(&nlm_blocked_lock);
126}
127
128/*
129 * Remove a block from the global list
130 */
131static inline void
132nlmsvc_remove_block(struct nlm_block *block)
133{
134 if (!list_empty(&block->b_list)) {
135 spin_lock(&nlm_blocked_lock);
136 list_del_init(&block->b_list);
137 spin_unlock(&nlm_blocked_lock);
138 nlmsvc_release_block(block);
139 }
140}
141
142/*
143 * Find a block for a given lock
144 */
145static struct nlm_block *
146nlmsvc_lookup_block(struct nlm_file *file, struct nlm_lock *lock)
147{
148 struct nlm_block *block;
149 struct file_lock *fl;
150
151 dprintk("lockd: nlmsvc_lookup_block f=%p pd=%d %Ld-%Ld ty=%d\n",
152 file, lock->fl.fl_pid,
153 (long long)lock->fl.fl_start,
154 (long long)lock->fl.fl_end, lock->fl.fl_type);
155 list_for_each_entry(block, &nlm_blocked, b_list) {
156 fl = &block->b_call->a_args.lock.fl;
157 dprintk("lockd: check f=%p pd=%d %Ld-%Ld ty=%d cookie=%s\n",
158 block->b_file, fl->fl_pid,
159 (long long)fl->fl_start,
160 (long long)fl->fl_end, fl->fl_type,
161 nlmdbg_cookie2a(&block->b_call->a_args.cookie));
162 if (block->b_file == file && nlm_compare_locks(fl, &lock->fl)) {
163 kref_get(&block->b_count);
164 return block;
165 }
166 }
167
168 return NULL;
169}
170
171static inline int nlm_cookie_match(struct nlm_cookie *a, struct nlm_cookie *b)
172{
173 if (a->len != b->len)
174 return 0;
175 if (memcmp(a->data, b->data, a->len))
176 return 0;
177 return 1;
178}
179
180/*
181 * Find a block with a given NLM cookie.
182 */
183static inline struct nlm_block *
184nlmsvc_find_block(struct nlm_cookie *cookie)
185{
186 struct nlm_block *block;
187
188 list_for_each_entry(block, &nlm_blocked, b_list) {
189 if (nlm_cookie_match(&block->b_call->a_args.cookie,cookie))
190 goto found;
191 }
192
193 return NULL;
194
195found:
196 dprintk("nlmsvc_find_block(%s): block=%p\n", nlmdbg_cookie2a(cookie), block);
197 kref_get(&block->b_count);
198 return block;
199}
200
201/*
202 * Create a block and initialize it.
203 *
204 * Note: we explicitly set the cookie of the grant reply to that of
205 * the blocked lock request. The spec explicitly mentions that the client
206 * should _not_ rely on the callback containing the same cookie as the
207 * request, but (as I found out later) that's because some implementations
208 * do just this. Never mind the standards comittees, they support our
209 * logging industries.
210 *
211 * 10 years later: I hope we can safely ignore these old and broken
212 * clients by now. Let's fix this so we can uniquely identify an incoming
213 * GRANTED_RES message by cookie, without having to rely on the client's IP
214 * address. --okir
215 */
216static struct nlm_block *
217nlmsvc_create_block(struct svc_rqst *rqstp, struct nlm_host *host,
218 struct nlm_file *file, struct nlm_lock *lock,
219 struct nlm_cookie *cookie)
220{
221 struct nlm_block *block;
222 struct nlm_rqst *call = NULL;
223
224 call = nlm_alloc_call(host);
225 if (call == NULL)
226 return NULL;
227
228 /* Allocate memory for block, and initialize arguments */
229 block = kzalloc(sizeof(*block), GFP_KERNEL);
230 if (block == NULL)
231 goto failed;
232 kref_init(&block->b_count);
233 INIT_LIST_HEAD(&block->b_list);
234 INIT_LIST_HEAD(&block->b_flist);
235
236 if (!nlmsvc_setgrantargs(call, lock))
237 goto failed_free;
238
239 /* Set notifier function for VFS, and init args */
240 call->a_args.lock.fl.fl_flags |= FL_SLEEP;
241 call->a_args.lock.fl.fl_lmops = &nlmsvc_lock_operations;
242 nlmclnt_next_cookie(&call->a_args.cookie);
243
244 dprintk("lockd: created block %p...\n", block);
245
246 /* Create and initialize the block */
247 block->b_daemon = rqstp->rq_server;
248 block->b_host = host;
249 block->b_file = file;
250 file->f_count++;
251
252 /* Add to file's list of blocks */
253 list_add(&block->b_flist, &file->f_blocks);
254
255 /* Set up RPC arguments for callback */
256 block->b_call = call;
257 call->a_flags = RPC_TASK_ASYNC;
258 call->a_block = block;
259
260 return block;
261
262failed_free:
263 kfree(block);
264failed:
265 nlmsvc_release_call(call);
266 return NULL;
267}
268
269/*
270 * Delete a block.
271 * It is the caller's responsibility to check whether the file
272 * can be closed hereafter.
273 */
274static int nlmsvc_unlink_block(struct nlm_block *block)
275{
276 int status;
277 dprintk("lockd: unlinking block %p...\n", block);
278
279 /* Remove block from list */
280 status = locks_delete_block(&block->b_call->a_args.lock.fl);
281 nlmsvc_remove_block(block);
282 return status;
283}
284
285static void nlmsvc_free_block(struct kref *kref)
286{
287 struct nlm_block *block = container_of(kref, struct nlm_block, b_count);
288 struct nlm_file *file = block->b_file;
289
290 dprintk("lockd: freeing block %p...\n", block);
291
292 /* Remove block from file's list of blocks */
293 list_del_init(&block->b_flist);
294 mutex_unlock(&file->f_mutex);
295
296 nlmsvc_freegrantargs(block->b_call);
297 nlmsvc_release_call(block->b_call);
298 nlm_release_file(block->b_file);
299 kfree(block);
300}
301
302static void nlmsvc_release_block(struct nlm_block *block)
303{
304 if (block != NULL)
305 kref_put_mutex(&block->b_count, nlmsvc_free_block, &block->b_file->f_mutex);
306}
307
308/*
309 * Loop over all blocks and delete blocks held by
310 * a matching host.
311 */
312void nlmsvc_traverse_blocks(struct nlm_host *host,
313 struct nlm_file *file,
314 nlm_host_match_fn_t match)
315{
316 struct nlm_block *block, *next;
317
318restart:
319 mutex_lock(&file->f_mutex);
320 list_for_each_entry_safe(block, next, &file->f_blocks, b_flist) {
321 if (!match(block->b_host, host))
322 continue;
323 /* Do not destroy blocks that are not on
324 * the global retry list - why? */
325 if (list_empty(&block->b_list))
326 continue;
327 kref_get(&block->b_count);
328 mutex_unlock(&file->f_mutex);
329 nlmsvc_unlink_block(block);
330 nlmsvc_release_block(block);
331 goto restart;
332 }
333 mutex_unlock(&file->f_mutex);
334}
335
336static struct nlm_lockowner *
337nlmsvc_get_lockowner(struct nlm_lockowner *lockowner)
338{
339 refcount_inc(&lockowner->count);
340 return lockowner;
341}
342
343void nlmsvc_put_lockowner(struct nlm_lockowner *lockowner)
344{
345 if (!refcount_dec_and_lock(&lockowner->count, &lockowner->host->h_lock))
346 return;
347 list_del(&lockowner->list);
348 spin_unlock(&lockowner->host->h_lock);
349 nlmsvc_release_host(lockowner->host);
350 kfree(lockowner);
351}
352
353static struct nlm_lockowner *__nlmsvc_find_lockowner(struct nlm_host *host, pid_t pid)
354{
355 struct nlm_lockowner *lockowner;
356 list_for_each_entry(lockowner, &host->h_lockowners, list) {
357 if (lockowner->pid != pid)
358 continue;
359 return nlmsvc_get_lockowner(lockowner);
360 }
361 return NULL;
362}
363
364static struct nlm_lockowner *nlmsvc_find_lockowner(struct nlm_host *host, pid_t pid)
365{
366 struct nlm_lockowner *res, *new = NULL;
367
368 spin_lock(&host->h_lock);
369 res = __nlmsvc_find_lockowner(host, pid);
370
371 if (res == NULL) {
372 spin_unlock(&host->h_lock);
373 new = kmalloc(sizeof(*res), GFP_KERNEL);
374 spin_lock(&host->h_lock);
375 res = __nlmsvc_find_lockowner(host, pid);
376 if (res == NULL && new != NULL) {
377 res = new;
378 /* fs/locks.c will manage the refcount through lock_ops */
379 refcount_set(&new->count, 1);
380 new->pid = pid;
381 new->host = nlm_get_host(host);
382 list_add(&new->list, &host->h_lockowners);
383 new = NULL;
384 }
385 }
386
387 spin_unlock(&host->h_lock);
388 kfree(new);
389 return res;
390}
391
392void
393nlmsvc_release_lockowner(struct nlm_lock *lock)
394{
395 if (lock->fl.fl_owner)
396 nlmsvc_put_lockowner(lock->fl.fl_owner);
397}
398
399void nlmsvc_locks_init_private(struct file_lock *fl, struct nlm_host *host,
400 pid_t pid)
401{
402 fl->fl_owner = nlmsvc_find_lockowner(host, pid);
403}
404
405/*
406 * Initialize arguments for GRANTED call. The nlm_rqst structure
407 * has been cleared already.
408 */
409static int nlmsvc_setgrantargs(struct nlm_rqst *call, struct nlm_lock *lock)
410{
411 locks_copy_lock(&call->a_args.lock.fl, &lock->fl);
412 memcpy(&call->a_args.lock.fh, &lock->fh, sizeof(call->a_args.lock.fh));
413 call->a_args.lock.caller = utsname()->nodename;
414 call->a_args.lock.oh.len = lock->oh.len;
415
416 /* set default data area */
417 call->a_args.lock.oh.data = call->a_owner;
418 call->a_args.lock.svid = ((struct nlm_lockowner *)lock->fl.fl_owner)->pid;
419
420 if (lock->oh.len > NLMCLNT_OHSIZE) {
421 void *data = kmalloc(lock->oh.len, GFP_KERNEL);
422 if (!data)
423 return 0;
424 call->a_args.lock.oh.data = (u8 *) data;
425 }
426
427 memcpy(call->a_args.lock.oh.data, lock->oh.data, lock->oh.len);
428 return 1;
429}
430
431static void nlmsvc_freegrantargs(struct nlm_rqst *call)
432{
433 if (call->a_args.lock.oh.data != call->a_owner)
434 kfree(call->a_args.lock.oh.data);
435
436 locks_release_private(&call->a_args.lock.fl);
437}
438
439/*
440 * Deferred lock request handling for non-blocking lock
441 */
442static __be32
443nlmsvc_defer_lock_rqst(struct svc_rqst *rqstp, struct nlm_block *block)
444{
445 __be32 status = nlm_lck_denied_nolocks;
446
447 block->b_flags |= B_QUEUED;
448
449 nlmsvc_insert_block(block, NLM_TIMEOUT);
450
451 block->b_cache_req = &rqstp->rq_chandle;
452 if (rqstp->rq_chandle.defer) {
453 block->b_deferred_req =
454 rqstp->rq_chandle.defer(block->b_cache_req);
455 if (block->b_deferred_req != NULL)
456 status = nlm_drop_reply;
457 }
458 dprintk("lockd: nlmsvc_defer_lock_rqst block %p flags %d status %d\n",
459 block, block->b_flags, ntohl(status));
460
461 return status;
462}
463
464/*
465 * Attempt to establish a lock, and if it can't be granted, block it
466 * if required.
467 */
468__be32
469nlmsvc_lock(struct svc_rqst *rqstp, struct nlm_file *file,
470 struct nlm_host *host, struct nlm_lock *lock, int wait,
471 struct nlm_cookie *cookie, int reclaim)
472{
473#if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
474 struct inode *inode = nlmsvc_file_inode(file);
475#endif
476 struct nlm_block *block = NULL;
477 int error;
478 int mode;
479 int async_block = 0;
480 __be32 ret;
481
482 dprintk("lockd: nlmsvc_lock(%s/%ld, ty=%d, pi=%d, %Ld-%Ld, bl=%d)\n",
483 inode->i_sb->s_id, inode->i_ino,
484 lock->fl.fl_type, lock->fl.fl_pid,
485 (long long)lock->fl.fl_start,
486 (long long)lock->fl.fl_end,
487 wait);
488
489 if (nlmsvc_file_file(file)->f_op->lock) {
490 async_block = wait;
491 wait = 0;
492 }
493
494 /* Lock file against concurrent access */
495 mutex_lock(&file->f_mutex);
496 /* Get existing block (in case client is busy-waiting)
497 * or create new block
498 */
499 block = nlmsvc_lookup_block(file, lock);
500 if (block == NULL) {
501 block = nlmsvc_create_block(rqstp, host, file, lock, cookie);
502 ret = nlm_lck_denied_nolocks;
503 if (block == NULL)
504 goto out;
505 lock = &block->b_call->a_args.lock;
506 } else
507 lock->fl.fl_flags &= ~FL_SLEEP;
508
509 if (block->b_flags & B_QUEUED) {
510 dprintk("lockd: nlmsvc_lock deferred block %p flags %d\n",
511 block, block->b_flags);
512 if (block->b_granted) {
513 nlmsvc_unlink_block(block);
514 ret = nlm_granted;
515 goto out;
516 }
517 if (block->b_flags & B_TIMED_OUT) {
518 nlmsvc_unlink_block(block);
519 ret = nlm_lck_denied;
520 goto out;
521 }
522 ret = nlm_drop_reply;
523 goto out;
524 }
525
526 if (locks_in_grace(SVC_NET(rqstp)) && !reclaim) {
527 ret = nlm_lck_denied_grace_period;
528 goto out;
529 }
530 if (reclaim && !locks_in_grace(SVC_NET(rqstp))) {
531 ret = nlm_lck_denied_grace_period;
532 goto out;
533 }
534
535 if (!wait)
536 lock->fl.fl_flags &= ~FL_SLEEP;
537 mode = lock_to_openmode(&lock->fl);
538 error = vfs_lock_file(file->f_file[mode], F_SETLK, &lock->fl, NULL);
539 lock->fl.fl_flags &= ~FL_SLEEP;
540
541 dprintk("lockd: vfs_lock_file returned %d\n", error);
542 switch (error) {
543 case 0:
544 ret = nlm_granted;
545 goto out;
546 case -EAGAIN:
547 /*
548 * If this is a blocking request for an
549 * already pending lock request then we need
550 * to put it back on lockd's block list
551 */
552 if (wait)
553 break;
554 ret = async_block ? nlm_lck_blocked : nlm_lck_denied;
555 goto out;
556 case FILE_LOCK_DEFERRED:
557 if (wait)
558 break;
559 /* Filesystem lock operation is in progress
560 Add it to the queue waiting for callback */
561 ret = nlmsvc_defer_lock_rqst(rqstp, block);
562 goto out;
563 case -EDEADLK:
564 ret = nlm_deadlock;
565 goto out;
566 default: /* includes ENOLCK */
567 ret = nlm_lck_denied_nolocks;
568 goto out;
569 }
570
571 ret = nlm_lck_blocked;
572
573 /* Append to list of blocked */
574 nlmsvc_insert_block(block, NLM_NEVER);
575out:
576 mutex_unlock(&file->f_mutex);
577 nlmsvc_release_block(block);
578 dprintk("lockd: nlmsvc_lock returned %u\n", ret);
579 return ret;
580}
581
582/*
583 * Test for presence of a conflicting lock.
584 */
585__be32
586nlmsvc_testlock(struct svc_rqst *rqstp, struct nlm_file *file,
587 struct nlm_host *host, struct nlm_lock *lock,
588 struct nlm_lock *conflock, struct nlm_cookie *cookie)
589{
590 int error;
591 int mode;
592 __be32 ret;
593
594 dprintk("lockd: nlmsvc_testlock(%s/%ld, ty=%d, %Ld-%Ld)\n",
595 nlmsvc_file_inode(file)->i_sb->s_id,
596 nlmsvc_file_inode(file)->i_ino,
597 lock->fl.fl_type,
598 (long long)lock->fl.fl_start,
599 (long long)lock->fl.fl_end);
600
601 if (locks_in_grace(SVC_NET(rqstp))) {
602 ret = nlm_lck_denied_grace_period;
603 goto out;
604 }
605
606 mode = lock_to_openmode(&lock->fl);
607 error = vfs_test_lock(file->f_file[mode], &lock->fl);
608 if (error) {
609 /* We can't currently deal with deferred test requests */
610 if (error == FILE_LOCK_DEFERRED)
611 WARN_ON_ONCE(1);
612
613 ret = nlm_lck_denied_nolocks;
614 goto out;
615 }
616
617 if (lock->fl.fl_type == F_UNLCK) {
618 ret = nlm_granted;
619 goto out;
620 }
621
622 dprintk("lockd: conflicting lock(ty=%d, %Ld-%Ld)\n",
623 lock->fl.fl_type, (long long)lock->fl.fl_start,
624 (long long)lock->fl.fl_end);
625 conflock->caller = "somehost"; /* FIXME */
626 conflock->len = strlen(conflock->caller);
627 conflock->oh.len = 0; /* don't return OH info */
628 conflock->svid = lock->fl.fl_pid;
629 conflock->fl.fl_type = lock->fl.fl_type;
630 conflock->fl.fl_start = lock->fl.fl_start;
631 conflock->fl.fl_end = lock->fl.fl_end;
632 locks_release_private(&lock->fl);
633
634 ret = nlm_lck_denied;
635out:
636 return ret;
637}
638
639/*
640 * Remove a lock.
641 * This implies a CANCEL call: We send a GRANT_MSG, the client replies
642 * with a GRANT_RES call which gets lost, and calls UNLOCK immediately
643 * afterwards. In this case the block will still be there, and hence
644 * must be removed.
645 */
646__be32
647nlmsvc_unlock(struct net *net, struct nlm_file *file, struct nlm_lock *lock)
648{
649 int error = 0;
650
651 dprintk("lockd: nlmsvc_unlock(%s/%ld, pi=%d, %Ld-%Ld)\n",
652 nlmsvc_file_inode(file)->i_sb->s_id,
653 nlmsvc_file_inode(file)->i_ino,
654 lock->fl.fl_pid,
655 (long long)lock->fl.fl_start,
656 (long long)lock->fl.fl_end);
657
658 /* First, cancel any lock that might be there */
659 nlmsvc_cancel_blocked(net, file, lock);
660
661 lock->fl.fl_type = F_UNLCK;
662 lock->fl.fl_file = file->f_file[O_RDONLY];
663 if (lock->fl.fl_file)
664 error = vfs_lock_file(lock->fl.fl_file, F_SETLK,
665 &lock->fl, NULL);
666 lock->fl.fl_file = file->f_file[O_WRONLY];
667 if (lock->fl.fl_file)
668 error |= vfs_lock_file(lock->fl.fl_file, F_SETLK,
669 &lock->fl, NULL);
670
671 return (error < 0)? nlm_lck_denied_nolocks : nlm_granted;
672}
673
674/*
675 * Cancel a previously blocked request.
676 *
677 * A cancel request always overrides any grant that may currently
678 * be in progress.
679 * The calling procedure must check whether the file can be closed.
680 */
681__be32
682nlmsvc_cancel_blocked(struct net *net, struct nlm_file *file, struct nlm_lock *lock)
683{
684 struct nlm_block *block;
685 int status = 0;
686 int mode;
687
688 dprintk("lockd: nlmsvc_cancel(%s/%ld, pi=%d, %Ld-%Ld)\n",
689 nlmsvc_file_inode(file)->i_sb->s_id,
690 nlmsvc_file_inode(file)->i_ino,
691 lock->fl.fl_pid,
692 (long long)lock->fl.fl_start,
693 (long long)lock->fl.fl_end);
694
695 if (locks_in_grace(net))
696 return nlm_lck_denied_grace_period;
697
698 mutex_lock(&file->f_mutex);
699 block = nlmsvc_lookup_block(file, lock);
700 mutex_unlock(&file->f_mutex);
701 if (block != NULL) {
702 struct file_lock *fl = &block->b_call->a_args.lock.fl;
703
704 mode = lock_to_openmode(fl);
705 vfs_cancel_lock(block->b_file->f_file[mode], fl);
706 status = nlmsvc_unlink_block(block);
707 nlmsvc_release_block(block);
708 }
709 return status ? nlm_lck_denied : nlm_granted;
710}
711
712/*
713 * This is a callback from the filesystem for VFS file lock requests.
714 * It will be used if lm_grant is defined and the filesystem can not
715 * respond to the request immediately.
716 * For SETLK or SETLKW request it will get the local posix lock.
717 * In all cases it will move the block to the head of nlm_blocked q where
718 * nlmsvc_retry_blocked() can send back a reply for SETLKW or revisit the
719 * deferred rpc for GETLK and SETLK.
720 */
721static void
722nlmsvc_update_deferred_block(struct nlm_block *block, int result)
723{
724 block->b_flags |= B_GOT_CALLBACK;
725 if (result == 0)
726 block->b_granted = 1;
727 else
728 block->b_flags |= B_TIMED_OUT;
729}
730
731static int nlmsvc_grant_deferred(struct file_lock *fl, int result)
732{
733 struct nlm_block *block;
734 int rc = -ENOENT;
735
736 spin_lock(&nlm_blocked_lock);
737 list_for_each_entry(block, &nlm_blocked, b_list) {
738 if (nlm_compare_locks(&block->b_call->a_args.lock.fl, fl)) {
739 dprintk("lockd: nlmsvc_notify_blocked block %p flags %d\n",
740 block, block->b_flags);
741 if (block->b_flags & B_QUEUED) {
742 if (block->b_flags & B_TIMED_OUT) {
743 rc = -ENOLCK;
744 break;
745 }
746 nlmsvc_update_deferred_block(block, result);
747 } else if (result == 0)
748 block->b_granted = 1;
749
750 nlmsvc_insert_block_locked(block, 0);
751 svc_wake_up(block->b_daemon);
752 rc = 0;
753 break;
754 }
755 }
756 spin_unlock(&nlm_blocked_lock);
757 if (rc == -ENOENT)
758 printk(KERN_WARNING "lockd: grant for unknown block\n");
759 return rc;
760}
761
762/*
763 * Unblock a blocked lock request. This is a callback invoked from the
764 * VFS layer when a lock on which we blocked is removed.
765 *
766 * This function doesn't grant the blocked lock instantly, but rather moves
767 * the block to the head of nlm_blocked where it can be picked up by lockd.
768 */
769static void
770nlmsvc_notify_blocked(struct file_lock *fl)
771{
772 struct nlm_block *block;
773
774 dprintk("lockd: VFS unblock notification for block %p\n", fl);
775 spin_lock(&nlm_blocked_lock);
776 list_for_each_entry(block, &nlm_blocked, b_list) {
777 if (nlm_compare_locks(&block->b_call->a_args.lock.fl, fl)) {
778 nlmsvc_insert_block_locked(block, 0);
779 spin_unlock(&nlm_blocked_lock);
780 svc_wake_up(block->b_daemon);
781 return;
782 }
783 }
784 spin_unlock(&nlm_blocked_lock);
785 printk(KERN_WARNING "lockd: notification for unknown block!\n");
786}
787
788static fl_owner_t nlmsvc_get_owner(fl_owner_t owner)
789{
790 return nlmsvc_get_lockowner(owner);
791}
792
793static void nlmsvc_put_owner(fl_owner_t owner)
794{
795 nlmsvc_put_lockowner(owner);
796}
797
798const struct lock_manager_operations nlmsvc_lock_operations = {
799 .lm_notify = nlmsvc_notify_blocked,
800 .lm_grant = nlmsvc_grant_deferred,
801 .lm_get_owner = nlmsvc_get_owner,
802 .lm_put_owner = nlmsvc_put_owner,
803};
804
805/*
806 * Try to claim a lock that was previously blocked.
807 *
808 * Note that we use both the RPC_GRANTED_MSG call _and_ an async
809 * RPC thread when notifying the client. This seems like overkill...
810 * Here's why:
811 * - we don't want to use a synchronous RPC thread, otherwise
812 * we might find ourselves hanging on a dead portmapper.
813 * - Some lockd implementations (e.g. HP) don't react to
814 * RPC_GRANTED calls; they seem to insist on RPC_GRANTED_MSG calls.
815 */
816static void
817nlmsvc_grant_blocked(struct nlm_block *block)
818{
819 struct nlm_file *file = block->b_file;
820 struct nlm_lock *lock = &block->b_call->a_args.lock;
821 int mode;
822 int error;
823 loff_t fl_start, fl_end;
824
825 dprintk("lockd: grant blocked lock %p\n", block);
826
827 kref_get(&block->b_count);
828
829 /* Unlink block request from list */
830 nlmsvc_unlink_block(block);
831
832 /* If b_granted is true this means we've been here before.
833 * Just retry the grant callback, possibly refreshing the RPC
834 * binding */
835 if (block->b_granted) {
836 nlm_rebind_host(block->b_host);
837 goto callback;
838 }
839
840 /* Try the lock operation again */
841 /* vfs_lock_file() can mangle fl_start and fl_end, but we need
842 * them unchanged for the GRANT_MSG
843 */
844 lock->fl.fl_flags |= FL_SLEEP;
845 fl_start = lock->fl.fl_start;
846 fl_end = lock->fl.fl_end;
847 mode = lock_to_openmode(&lock->fl);
848 error = vfs_lock_file(file->f_file[mode], F_SETLK, &lock->fl, NULL);
849 lock->fl.fl_flags &= ~FL_SLEEP;
850 lock->fl.fl_start = fl_start;
851 lock->fl.fl_end = fl_end;
852
853 switch (error) {
854 case 0:
855 break;
856 case FILE_LOCK_DEFERRED:
857 dprintk("lockd: lock still blocked error %d\n", error);
858 nlmsvc_insert_block(block, NLM_NEVER);
859 nlmsvc_release_block(block);
860 return;
861 default:
862 printk(KERN_WARNING "lockd: unexpected error %d in %s!\n",
863 -error, __func__);
864 nlmsvc_insert_block(block, 10 * HZ);
865 nlmsvc_release_block(block);
866 return;
867 }
868
869callback:
870 /* Lock was granted by VFS. */
871 dprintk("lockd: GRANTing blocked lock.\n");
872 block->b_granted = 1;
873
874 /* keep block on the list, but don't reattempt until the RPC
875 * completes or the submission fails
876 */
877 nlmsvc_insert_block(block, NLM_NEVER);
878
879 /* Call the client -- use a soft RPC task since nlmsvc_retry_blocked
880 * will queue up a new one if this one times out
881 */
882 error = nlm_async_call(block->b_call, NLMPROC_GRANTED_MSG,
883 &nlmsvc_grant_ops);
884
885 /* RPC submission failed, wait a bit and retry */
886 if (error < 0)
887 nlmsvc_insert_block(block, 10 * HZ);
888}
889
890/*
891 * This is the callback from the RPC layer when the NLM_GRANTED_MSG
892 * RPC call has succeeded or timed out.
893 * Like all RPC callbacks, it is invoked by the rpciod process, so it
894 * better not sleep. Therefore, we put the blocked lock on the nlm_blocked
895 * chain once more in order to have it removed by lockd itself (which can
896 * then sleep on the file semaphore without disrupting e.g. the nfs client).
897 */
898static void nlmsvc_grant_callback(struct rpc_task *task, void *data)
899{
900 struct nlm_rqst *call = data;
901 struct nlm_block *block = call->a_block;
902 unsigned long timeout;
903
904 dprintk("lockd: GRANT_MSG RPC callback\n");
905
906 spin_lock(&nlm_blocked_lock);
907 /* if the block is not on a list at this point then it has
908 * been invalidated. Don't try to requeue it.
909 *
910 * FIXME: it's possible that the block is removed from the list
911 * after this check but before the nlmsvc_insert_block. In that
912 * case it will be added back. Perhaps we need better locking
913 * for nlm_blocked?
914 */
915 if (list_empty(&block->b_list))
916 goto out;
917
918 /* Technically, we should down the file semaphore here. Since we
919 * move the block towards the head of the queue only, no harm
920 * can be done, though. */
921 if (task->tk_status < 0) {
922 /* RPC error: Re-insert for retransmission */
923 timeout = 10 * HZ;
924 } else {
925 /* Call was successful, now wait for client callback */
926 timeout = 60 * HZ;
927 }
928 nlmsvc_insert_block_locked(block, timeout);
929 svc_wake_up(block->b_daemon);
930out:
931 spin_unlock(&nlm_blocked_lock);
932}
933
934/*
935 * FIXME: nlmsvc_release_block() grabs a mutex. This is not allowed for an
936 * .rpc_release rpc_call_op
937 */
938static void nlmsvc_grant_release(void *data)
939{
940 struct nlm_rqst *call = data;
941 nlmsvc_release_block(call->a_block);
942}
943
944static const struct rpc_call_ops nlmsvc_grant_ops = {
945 .rpc_call_done = nlmsvc_grant_callback,
946 .rpc_release = nlmsvc_grant_release,
947};
948
949/*
950 * We received a GRANT_RES callback. Try to find the corresponding
951 * block.
952 */
953void
954nlmsvc_grant_reply(struct nlm_cookie *cookie, __be32 status)
955{
956 struct nlm_block *block;
957
958 dprintk("grant_reply: looking for cookie %x, s=%d \n",
959 *(unsigned int *)(cookie->data), status);
960 if (!(block = nlmsvc_find_block(cookie)))
961 return;
962
963 if (status == nlm_lck_denied_grace_period) {
964 /* Try again in a couple of seconds */
965 nlmsvc_insert_block(block, 10 * HZ);
966 } else {
967 /*
968 * Lock is now held by client, or has been rejected.
969 * In both cases, the block should be removed.
970 */
971 nlmsvc_unlink_block(block);
972 }
973 nlmsvc_release_block(block);
974}
975
976/* Helper function to handle retry of a deferred block.
977 * If it is a blocking lock, call grant_blocked.
978 * For a non-blocking lock or test lock, revisit the request.
979 */
980static void
981retry_deferred_block(struct nlm_block *block)
982{
983 if (!(block->b_flags & B_GOT_CALLBACK))
984 block->b_flags |= B_TIMED_OUT;
985 nlmsvc_insert_block(block, NLM_TIMEOUT);
986 dprintk("revisit block %p flags %d\n", block, block->b_flags);
987 if (block->b_deferred_req) {
988 block->b_deferred_req->revisit(block->b_deferred_req, 0);
989 block->b_deferred_req = NULL;
990 }
991}
992
993/*
994 * Retry all blocked locks that have been notified. This is where lockd
995 * picks up locks that can be granted, or grant notifications that must
996 * be retransmitted.
997 */
998unsigned long
999nlmsvc_retry_blocked(void)
1000{
1001 unsigned long timeout = MAX_SCHEDULE_TIMEOUT;
1002 struct nlm_block *block;
1003
1004 spin_lock(&nlm_blocked_lock);
1005 while (!list_empty(&nlm_blocked) && !kthread_should_stop()) {
1006 block = list_entry(nlm_blocked.next, struct nlm_block, b_list);
1007
1008 if (block->b_when == NLM_NEVER)
1009 break;
1010 if (time_after(block->b_when, jiffies)) {
1011 timeout = block->b_when - jiffies;
1012 break;
1013 }
1014 spin_unlock(&nlm_blocked_lock);
1015
1016 dprintk("nlmsvc_retry_blocked(%p, when=%ld)\n",
1017 block, block->b_when);
1018 if (block->b_flags & B_QUEUED) {
1019 dprintk("nlmsvc_retry_blocked delete block (%p, granted=%d, flags=%d)\n",
1020 block, block->b_granted, block->b_flags);
1021 retry_deferred_block(block);
1022 } else
1023 nlmsvc_grant_blocked(block);
1024 spin_lock(&nlm_blocked_lock);
1025 }
1026 spin_unlock(&nlm_blocked_lock);
1027
1028 return timeout;
1029}
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#ifdef LOCKD_DEBUG
57static const char *nlmdbg_cookie2a(const struct nlm_cookie *cookie)
58{
59 /*
60 * We can get away with a static buffer because we're only
61 * called with BKL held.
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 block->b_fl = NULL;
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->b_fl);
299 kfree(block);
300}
301
302static void nlmsvc_release_block(struct nlm_block *block)
303{
304 if (block != NULL)
305 kref_put_mutex(&block->b_count, nlmsvc_free_block, &block->b_file->f_mutex);
306}
307
308/*
309 * Loop over all blocks and delete blocks held by
310 * a matching host.
311 */
312void nlmsvc_traverse_blocks(struct nlm_host *host,
313 struct nlm_file *file,
314 nlm_host_match_fn_t match)
315{
316 struct nlm_block *block, *next;
317
318restart:
319 mutex_lock(&file->f_mutex);
320 list_for_each_entry_safe(block, next, &file->f_blocks, b_flist) {
321 if (!match(block->b_host, host))
322 continue;
323 /* Do not destroy blocks that are not on
324 * the global retry list - why? */
325 if (list_empty(&block->b_list))
326 continue;
327 kref_get(&block->b_count);
328 mutex_unlock(&file->f_mutex);
329 nlmsvc_unlink_block(block);
330 nlmsvc_release_block(block);
331 goto restart;
332 }
333 mutex_unlock(&file->f_mutex);
334}
335
336/*
337 * Initialize arguments for GRANTED call. The nlm_rqst structure
338 * has been cleared already.
339 */
340static int nlmsvc_setgrantargs(struct nlm_rqst *call, struct nlm_lock *lock)
341{
342 locks_copy_lock(&call->a_args.lock.fl, &lock->fl);
343 memcpy(&call->a_args.lock.fh, &lock->fh, sizeof(call->a_args.lock.fh));
344 call->a_args.lock.caller = utsname()->nodename;
345 call->a_args.lock.oh.len = lock->oh.len;
346
347 /* set default data area */
348 call->a_args.lock.oh.data = call->a_owner;
349 call->a_args.lock.svid = lock->fl.fl_pid;
350
351 if (lock->oh.len > NLMCLNT_OHSIZE) {
352 void *data = kmalloc(lock->oh.len, GFP_KERNEL);
353 if (!data)
354 return 0;
355 call->a_args.lock.oh.data = (u8 *) data;
356 }
357
358 memcpy(call->a_args.lock.oh.data, lock->oh.data, lock->oh.len);
359 return 1;
360}
361
362static void nlmsvc_freegrantargs(struct nlm_rqst *call)
363{
364 if (call->a_args.lock.oh.data != call->a_owner)
365 kfree(call->a_args.lock.oh.data);
366
367 locks_release_private(&call->a_args.lock.fl);
368}
369
370/*
371 * Deferred lock request handling for non-blocking lock
372 */
373static __be32
374nlmsvc_defer_lock_rqst(struct svc_rqst *rqstp, struct nlm_block *block)
375{
376 __be32 status = nlm_lck_denied_nolocks;
377
378 block->b_flags |= B_QUEUED;
379
380 nlmsvc_insert_block(block, NLM_TIMEOUT);
381
382 block->b_cache_req = &rqstp->rq_chandle;
383 if (rqstp->rq_chandle.defer) {
384 block->b_deferred_req =
385 rqstp->rq_chandle.defer(block->b_cache_req);
386 if (block->b_deferred_req != NULL)
387 status = nlm_drop_reply;
388 }
389 dprintk("lockd: nlmsvc_defer_lock_rqst block %p flags %d status %d\n",
390 block, block->b_flags, ntohl(status));
391
392 return status;
393}
394
395/*
396 * Attempt to establish a lock, and if it can't be granted, block it
397 * if required.
398 */
399__be32
400nlmsvc_lock(struct svc_rqst *rqstp, struct nlm_file *file,
401 struct nlm_host *host, struct nlm_lock *lock, int wait,
402 struct nlm_cookie *cookie, int reclaim)
403{
404 struct nlm_block *block = NULL;
405 int error;
406 __be32 ret;
407
408 dprintk("lockd: nlmsvc_lock(%s/%ld, ty=%d, pi=%d, %Ld-%Ld, bl=%d)\n",
409 file_inode(file->f_file)->i_sb->s_id,
410 file_inode(file->f_file)->i_ino,
411 lock->fl.fl_type, lock->fl.fl_pid,
412 (long long)lock->fl.fl_start,
413 (long long)lock->fl.fl_end,
414 wait);
415
416 /* Lock file against concurrent access */
417 mutex_lock(&file->f_mutex);
418 /* Get existing block (in case client is busy-waiting)
419 * or create new block
420 */
421 block = nlmsvc_lookup_block(file, lock);
422 if (block == NULL) {
423 block = nlmsvc_create_block(rqstp, host, file, lock, cookie);
424 ret = nlm_lck_denied_nolocks;
425 if (block == NULL)
426 goto out;
427 lock = &block->b_call->a_args.lock;
428 } else
429 lock->fl.fl_flags &= ~FL_SLEEP;
430
431 if (block->b_flags & B_QUEUED) {
432 dprintk("lockd: nlmsvc_lock deferred block %p flags %d\n",
433 block, block->b_flags);
434 if (block->b_granted) {
435 nlmsvc_unlink_block(block);
436 ret = nlm_granted;
437 goto out;
438 }
439 if (block->b_flags & B_TIMED_OUT) {
440 nlmsvc_unlink_block(block);
441 ret = nlm_lck_denied;
442 goto out;
443 }
444 ret = nlm_drop_reply;
445 goto out;
446 }
447
448 if (locks_in_grace(SVC_NET(rqstp)) && !reclaim) {
449 ret = nlm_lck_denied_grace_period;
450 goto out;
451 }
452 if (reclaim && !locks_in_grace(SVC_NET(rqstp))) {
453 ret = nlm_lck_denied_grace_period;
454 goto out;
455 }
456
457 if (!wait)
458 lock->fl.fl_flags &= ~FL_SLEEP;
459 error = vfs_lock_file(file->f_file, F_SETLK, &lock->fl, NULL);
460 lock->fl.fl_flags &= ~FL_SLEEP;
461
462 dprintk("lockd: vfs_lock_file returned %d\n", error);
463 switch (error) {
464 case 0:
465 ret = nlm_granted;
466 goto out;
467 case -EAGAIN:
468 /*
469 * If this is a blocking request for an
470 * already pending lock request then we need
471 * to put it back on lockd's block list
472 */
473 if (wait)
474 break;
475 ret = nlm_lck_denied;
476 goto out;
477 case FILE_LOCK_DEFERRED:
478 if (wait)
479 break;
480 /* Filesystem lock operation is in progress
481 Add it to the queue waiting for callback */
482 ret = nlmsvc_defer_lock_rqst(rqstp, block);
483 goto out;
484 case -EDEADLK:
485 ret = nlm_deadlock;
486 goto out;
487 default: /* includes ENOLCK */
488 ret = nlm_lck_denied_nolocks;
489 goto out;
490 }
491
492 ret = nlm_lck_blocked;
493
494 /* Append to list of blocked */
495 nlmsvc_insert_block(block, NLM_NEVER);
496out:
497 mutex_unlock(&file->f_mutex);
498 nlmsvc_release_block(block);
499 dprintk("lockd: nlmsvc_lock returned %u\n", ret);
500 return ret;
501}
502
503/*
504 * Test for presence of a conflicting lock.
505 */
506__be32
507nlmsvc_testlock(struct svc_rqst *rqstp, struct nlm_file *file,
508 struct nlm_host *host, struct nlm_lock *lock,
509 struct nlm_lock *conflock, struct nlm_cookie *cookie)
510{
511 struct nlm_block *block = NULL;
512 int error;
513 __be32 ret;
514
515 dprintk("lockd: nlmsvc_testlock(%s/%ld, ty=%d, %Ld-%Ld)\n",
516 file_inode(file->f_file)->i_sb->s_id,
517 file_inode(file->f_file)->i_ino,
518 lock->fl.fl_type,
519 (long long)lock->fl.fl_start,
520 (long long)lock->fl.fl_end);
521
522 /* Get existing block (in case client is busy-waiting) */
523 block = nlmsvc_lookup_block(file, lock);
524
525 if (block == NULL) {
526 struct file_lock *conf = kzalloc(sizeof(*conf), GFP_KERNEL);
527
528 if (conf == NULL)
529 return nlm_granted;
530 block = nlmsvc_create_block(rqstp, host, file, lock, cookie);
531 if (block == NULL) {
532 kfree(conf);
533 return nlm_granted;
534 }
535 block->b_fl = conf;
536 }
537 if (block->b_flags & B_QUEUED) {
538 dprintk("lockd: nlmsvc_testlock deferred block %p flags %d fl %p\n",
539 block, block->b_flags, block->b_fl);
540 if (block->b_flags & B_TIMED_OUT) {
541 nlmsvc_unlink_block(block);
542 ret = nlm_lck_denied;
543 goto out;
544 }
545 if (block->b_flags & B_GOT_CALLBACK) {
546 nlmsvc_unlink_block(block);
547 if (block->b_fl != NULL
548 && block->b_fl->fl_type != F_UNLCK) {
549 lock->fl = *block->b_fl;
550 goto conf_lock;
551 } else {
552 ret = nlm_granted;
553 goto out;
554 }
555 }
556 ret = nlm_drop_reply;
557 goto out;
558 }
559
560 if (locks_in_grace(SVC_NET(rqstp))) {
561 ret = nlm_lck_denied_grace_period;
562 goto out;
563 }
564 error = vfs_test_lock(file->f_file, &lock->fl);
565 if (error == FILE_LOCK_DEFERRED) {
566 ret = nlmsvc_defer_lock_rqst(rqstp, block);
567 goto out;
568 }
569 if (error) {
570 ret = nlm_lck_denied_nolocks;
571 goto out;
572 }
573 if (lock->fl.fl_type == F_UNLCK) {
574 ret = nlm_granted;
575 goto out;
576 }
577
578conf_lock:
579 dprintk("lockd: conflicting lock(ty=%d, %Ld-%Ld)\n",
580 lock->fl.fl_type, (long long)lock->fl.fl_start,
581 (long long)lock->fl.fl_end);
582 conflock->caller = "somehost"; /* FIXME */
583 conflock->len = strlen(conflock->caller);
584 conflock->oh.len = 0; /* don't return OH info */
585 conflock->svid = lock->fl.fl_pid;
586 conflock->fl.fl_type = lock->fl.fl_type;
587 conflock->fl.fl_start = lock->fl.fl_start;
588 conflock->fl.fl_end = lock->fl.fl_end;
589 ret = nlm_lck_denied;
590out:
591 if (block)
592 nlmsvc_release_block(block);
593 return ret;
594}
595
596/*
597 * Remove a lock.
598 * This implies a CANCEL call: We send a GRANT_MSG, the client replies
599 * with a GRANT_RES call which gets lost, and calls UNLOCK immediately
600 * afterwards. In this case the block will still be there, and hence
601 * must be removed.
602 */
603__be32
604nlmsvc_unlock(struct net *net, struct nlm_file *file, struct nlm_lock *lock)
605{
606 int error;
607
608 dprintk("lockd: nlmsvc_unlock(%s/%ld, pi=%d, %Ld-%Ld)\n",
609 file_inode(file->f_file)->i_sb->s_id,
610 file_inode(file->f_file)->i_ino,
611 lock->fl.fl_pid,
612 (long long)lock->fl.fl_start,
613 (long long)lock->fl.fl_end);
614
615 /* First, cancel any lock that might be there */
616 nlmsvc_cancel_blocked(net, file, lock);
617
618 lock->fl.fl_type = F_UNLCK;
619 error = vfs_lock_file(file->f_file, F_SETLK, &lock->fl, NULL);
620
621 return (error < 0)? nlm_lck_denied_nolocks : nlm_granted;
622}
623
624/*
625 * Cancel a previously blocked request.
626 *
627 * A cancel request always overrides any grant that may currently
628 * be in progress.
629 * The calling procedure must check whether the file can be closed.
630 */
631__be32
632nlmsvc_cancel_blocked(struct net *net, struct nlm_file *file, struct nlm_lock *lock)
633{
634 struct nlm_block *block;
635 int status = 0;
636
637 dprintk("lockd: nlmsvc_cancel(%s/%ld, pi=%d, %Ld-%Ld)\n",
638 file_inode(file->f_file)->i_sb->s_id,
639 file_inode(file->f_file)->i_ino,
640 lock->fl.fl_pid,
641 (long long)lock->fl.fl_start,
642 (long long)lock->fl.fl_end);
643
644 if (locks_in_grace(net))
645 return nlm_lck_denied_grace_period;
646
647 mutex_lock(&file->f_mutex);
648 block = nlmsvc_lookup_block(file, lock);
649 mutex_unlock(&file->f_mutex);
650 if (block != NULL) {
651 vfs_cancel_lock(block->b_file->f_file,
652 &block->b_call->a_args.lock.fl);
653 status = nlmsvc_unlink_block(block);
654 nlmsvc_release_block(block);
655 }
656 return status ? nlm_lck_denied : nlm_granted;
657}
658
659/*
660 * This is a callback from the filesystem for VFS file lock requests.
661 * It will be used if lm_grant is defined and the filesystem can not
662 * respond to the request immediately.
663 * For GETLK request it will copy the reply to the nlm_block.
664 * For SETLK or SETLKW request it will get the local posix lock.
665 * In all cases it will move the block to the head of nlm_blocked q where
666 * nlmsvc_retry_blocked() can send back a reply for SETLKW or revisit the
667 * deferred rpc for GETLK and SETLK.
668 */
669static void
670nlmsvc_update_deferred_block(struct nlm_block *block, struct file_lock *conf,
671 int result)
672{
673 block->b_flags |= B_GOT_CALLBACK;
674 if (result == 0)
675 block->b_granted = 1;
676 else
677 block->b_flags |= B_TIMED_OUT;
678 if (conf) {
679 if (block->b_fl)
680 __locks_copy_lock(block->b_fl, conf);
681 }
682}
683
684static int nlmsvc_grant_deferred(struct file_lock *fl, struct file_lock *conf,
685 int result)
686{
687 struct nlm_block *block;
688 int rc = -ENOENT;
689
690 spin_lock(&nlm_blocked_lock);
691 list_for_each_entry(block, &nlm_blocked, b_list) {
692 if (nlm_compare_locks(&block->b_call->a_args.lock.fl, fl)) {
693 dprintk("lockd: nlmsvc_notify_blocked block %p flags %d\n",
694 block, block->b_flags);
695 if (block->b_flags & B_QUEUED) {
696 if (block->b_flags & B_TIMED_OUT) {
697 rc = -ENOLCK;
698 break;
699 }
700 nlmsvc_update_deferred_block(block, conf, result);
701 } else if (result == 0)
702 block->b_granted = 1;
703
704 nlmsvc_insert_block_locked(block, 0);
705 svc_wake_up(block->b_daemon);
706 rc = 0;
707 break;
708 }
709 }
710 spin_unlock(&nlm_blocked_lock);
711 if (rc == -ENOENT)
712 printk(KERN_WARNING "lockd: grant for unknown block\n");
713 return rc;
714}
715
716/*
717 * Unblock a blocked lock request. This is a callback invoked from the
718 * VFS layer when a lock on which we blocked is removed.
719 *
720 * This function doesn't grant the blocked lock instantly, but rather moves
721 * the block to the head of nlm_blocked where it can be picked up by lockd.
722 */
723static void
724nlmsvc_notify_blocked(struct file_lock *fl)
725{
726 struct nlm_block *block;
727
728 dprintk("lockd: VFS unblock notification for block %p\n", fl);
729 spin_lock(&nlm_blocked_lock);
730 list_for_each_entry(block, &nlm_blocked, b_list) {
731 if (nlm_compare_locks(&block->b_call->a_args.lock.fl, fl)) {
732 nlmsvc_insert_block_locked(block, 0);
733 spin_unlock(&nlm_blocked_lock);
734 svc_wake_up(block->b_daemon);
735 return;
736 }
737 }
738 spin_unlock(&nlm_blocked_lock);
739 printk(KERN_WARNING "lockd: notification for unknown block!\n");
740}
741
742static int nlmsvc_same_owner(struct file_lock *fl1, struct file_lock *fl2)
743{
744 return fl1->fl_owner == fl2->fl_owner && fl1->fl_pid == fl2->fl_pid;
745}
746
747/*
748 * Since NLM uses two "keys" for tracking locks, we need to hash them down
749 * to one for the blocked_hash. Here, we're just xor'ing the host address
750 * with the pid in order to create a key value for picking a hash bucket.
751 */
752static unsigned long
753nlmsvc_owner_key(struct file_lock *fl)
754{
755 return (unsigned long)fl->fl_owner ^ (unsigned long)fl->fl_pid;
756}
757
758const struct lock_manager_operations nlmsvc_lock_operations = {
759 .lm_compare_owner = nlmsvc_same_owner,
760 .lm_owner_key = nlmsvc_owner_key,
761 .lm_notify = nlmsvc_notify_blocked,
762 .lm_grant = nlmsvc_grant_deferred,
763};
764
765/*
766 * Try to claim a lock that was previously blocked.
767 *
768 * Note that we use both the RPC_GRANTED_MSG call _and_ an async
769 * RPC thread when notifying the client. This seems like overkill...
770 * Here's why:
771 * - we don't want to use a synchronous RPC thread, otherwise
772 * we might find ourselves hanging on a dead portmapper.
773 * - Some lockd implementations (e.g. HP) don't react to
774 * RPC_GRANTED calls; they seem to insist on RPC_GRANTED_MSG calls.
775 */
776static void
777nlmsvc_grant_blocked(struct nlm_block *block)
778{
779 struct nlm_file *file = block->b_file;
780 struct nlm_lock *lock = &block->b_call->a_args.lock;
781 int error;
782 loff_t fl_start, fl_end;
783
784 dprintk("lockd: grant blocked lock %p\n", block);
785
786 kref_get(&block->b_count);
787
788 /* Unlink block request from list */
789 nlmsvc_unlink_block(block);
790
791 /* If b_granted is true this means we've been here before.
792 * Just retry the grant callback, possibly refreshing the RPC
793 * binding */
794 if (block->b_granted) {
795 nlm_rebind_host(block->b_host);
796 goto callback;
797 }
798
799 /* Try the lock operation again */
800 /* vfs_lock_file() can mangle fl_start and fl_end, but we need
801 * them unchanged for the GRANT_MSG
802 */
803 lock->fl.fl_flags |= FL_SLEEP;
804 fl_start = lock->fl.fl_start;
805 fl_end = lock->fl.fl_end;
806 error = vfs_lock_file(file->f_file, F_SETLK, &lock->fl, NULL);
807 lock->fl.fl_flags &= ~FL_SLEEP;
808 lock->fl.fl_start = fl_start;
809 lock->fl.fl_end = fl_end;
810
811 switch (error) {
812 case 0:
813 break;
814 case FILE_LOCK_DEFERRED:
815 dprintk("lockd: lock still blocked error %d\n", error);
816 nlmsvc_insert_block(block, NLM_NEVER);
817 nlmsvc_release_block(block);
818 return;
819 default:
820 printk(KERN_WARNING "lockd: unexpected error %d in %s!\n",
821 -error, __func__);
822 nlmsvc_insert_block(block, 10 * HZ);
823 nlmsvc_release_block(block);
824 return;
825 }
826
827callback:
828 /* Lock was granted by VFS. */
829 dprintk("lockd: GRANTing blocked lock.\n");
830 block->b_granted = 1;
831
832 /* keep block on the list, but don't reattempt until the RPC
833 * completes or the submission fails
834 */
835 nlmsvc_insert_block(block, NLM_NEVER);
836
837 /* Call the client -- use a soft RPC task since nlmsvc_retry_blocked
838 * will queue up a new one if this one times out
839 */
840 error = nlm_async_call(block->b_call, NLMPROC_GRANTED_MSG,
841 &nlmsvc_grant_ops);
842
843 /* RPC submission failed, wait a bit and retry */
844 if (error < 0)
845 nlmsvc_insert_block(block, 10 * HZ);
846}
847
848/*
849 * This is the callback from the RPC layer when the NLM_GRANTED_MSG
850 * RPC call has succeeded or timed out.
851 * Like all RPC callbacks, it is invoked by the rpciod process, so it
852 * better not sleep. Therefore, we put the blocked lock on the nlm_blocked
853 * chain once more in order to have it removed by lockd itself (which can
854 * then sleep on the file semaphore without disrupting e.g. the nfs client).
855 */
856static void nlmsvc_grant_callback(struct rpc_task *task, void *data)
857{
858 struct nlm_rqst *call = data;
859 struct nlm_block *block = call->a_block;
860 unsigned long timeout;
861
862 dprintk("lockd: GRANT_MSG RPC callback\n");
863
864 spin_lock(&nlm_blocked_lock);
865 /* if the block is not on a list at this point then it has
866 * been invalidated. Don't try to requeue it.
867 *
868 * FIXME: it's possible that the block is removed from the list
869 * after this check but before the nlmsvc_insert_block. In that
870 * case it will be added back. Perhaps we need better locking
871 * for nlm_blocked?
872 */
873 if (list_empty(&block->b_list))
874 goto out;
875
876 /* Technically, we should down the file semaphore here. Since we
877 * move the block towards the head of the queue only, no harm
878 * can be done, though. */
879 if (task->tk_status < 0) {
880 /* RPC error: Re-insert for retransmission */
881 timeout = 10 * HZ;
882 } else {
883 /* Call was successful, now wait for client callback */
884 timeout = 60 * HZ;
885 }
886 nlmsvc_insert_block_locked(block, timeout);
887 svc_wake_up(block->b_daemon);
888out:
889 spin_unlock(&nlm_blocked_lock);
890}
891
892/*
893 * FIXME: nlmsvc_release_block() grabs a mutex. This is not allowed for an
894 * .rpc_release rpc_call_op
895 */
896static void nlmsvc_grant_release(void *data)
897{
898 struct nlm_rqst *call = data;
899 nlmsvc_release_block(call->a_block);
900}
901
902static const struct rpc_call_ops nlmsvc_grant_ops = {
903 .rpc_call_done = nlmsvc_grant_callback,
904 .rpc_release = nlmsvc_grant_release,
905};
906
907/*
908 * We received a GRANT_RES callback. Try to find the corresponding
909 * block.
910 */
911void
912nlmsvc_grant_reply(struct nlm_cookie *cookie, __be32 status)
913{
914 struct nlm_block *block;
915
916 dprintk("grant_reply: looking for cookie %x, s=%d \n",
917 *(unsigned int *)(cookie->data), status);
918 if (!(block = nlmsvc_find_block(cookie)))
919 return;
920
921 if (block) {
922 if (status == nlm_lck_denied_grace_period) {
923 /* Try again in a couple of seconds */
924 nlmsvc_insert_block(block, 10 * HZ);
925 } else {
926 /* Lock is now held by client, or has been rejected.
927 * In both cases, the block should be removed. */
928 nlmsvc_unlink_block(block);
929 }
930 }
931 nlmsvc_release_block(block);
932}
933
934/* Helper function to handle retry of a deferred block.
935 * If it is a blocking lock, call grant_blocked.
936 * For a non-blocking lock or test lock, revisit the request.
937 */
938static void
939retry_deferred_block(struct nlm_block *block)
940{
941 if (!(block->b_flags & B_GOT_CALLBACK))
942 block->b_flags |= B_TIMED_OUT;
943 nlmsvc_insert_block(block, NLM_TIMEOUT);
944 dprintk("revisit block %p flags %d\n", block, block->b_flags);
945 if (block->b_deferred_req) {
946 block->b_deferred_req->revisit(block->b_deferred_req, 0);
947 block->b_deferred_req = NULL;
948 }
949}
950
951/*
952 * Retry all blocked locks that have been notified. This is where lockd
953 * picks up locks that can be granted, or grant notifications that must
954 * be retransmitted.
955 */
956unsigned long
957nlmsvc_retry_blocked(void)
958{
959 unsigned long timeout = MAX_SCHEDULE_TIMEOUT;
960 struct nlm_block *block;
961
962 spin_lock(&nlm_blocked_lock);
963 while (!list_empty(&nlm_blocked) && !kthread_should_stop()) {
964 block = list_entry(nlm_blocked.next, struct nlm_block, b_list);
965
966 if (block->b_when == NLM_NEVER)
967 break;
968 if (time_after(block->b_when, jiffies)) {
969 timeout = block->b_when - jiffies;
970 break;
971 }
972 spin_unlock(&nlm_blocked_lock);
973
974 dprintk("nlmsvc_retry_blocked(%p, when=%ld)\n",
975 block, block->b_when);
976 if (block->b_flags & B_QUEUED) {
977 dprintk("nlmsvc_retry_blocked delete block (%p, granted=%d, flags=%d)\n",
978 block, block->b_granted, block->b_flags);
979 retry_deferred_block(block);
980 } else
981 nlmsvc_grant_blocked(block);
982 spin_lock(&nlm_blocked_lock);
983 }
984 spin_unlock(&nlm_blocked_lock);
985
986 return timeout;
987}