Merge branch 'fb' into devel
[linux-2.6] / fs / lockd / svclock.c
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/errno.h>
25 #include <linux/kernel.h>
26 #include <linux/sched.h>
27 #include <linux/smp_lock.h>
28 #include <linux/sunrpc/clnt.h>
29 #include <linux/sunrpc/svc.h>
30 #include <linux/lockd/nlm.h>
31 #include <linux/lockd/lockd.h>
32 #include <linux/kthread.h>
33
34 #define NLMDBG_FACILITY         NLMDBG_SVCLOCK
35
36 #ifdef CONFIG_LOCKD_V4
37 #define nlm_deadlock    nlm4_deadlock
38 #else
39 #define nlm_deadlock    nlm_lck_denied
40 #endif
41
42 static void nlmsvc_release_block(struct nlm_block *block);
43 static void     nlmsvc_insert_block(struct nlm_block *block, unsigned long);
44 static void     nlmsvc_remove_block(struct nlm_block *block);
45
46 static int nlmsvc_setgrantargs(struct nlm_rqst *call, struct nlm_lock *lock);
47 static void nlmsvc_freegrantargs(struct nlm_rqst *call);
48 static const struct rpc_call_ops nlmsvc_grant_ops;
49
50 /*
51  * The list of blocked locks to retry
52  */
53 static LIST_HEAD(nlm_blocked);
54
55 /*
56  * Insert a blocked lock into the global list
57  */
58 static void
59 nlmsvc_insert_block(struct nlm_block *block, unsigned long when)
60 {
61         struct nlm_block *b;
62         struct list_head *pos;
63
64         dprintk("lockd: nlmsvc_insert_block(%p, %ld)\n", block, when);
65         if (list_empty(&block->b_list)) {
66                 kref_get(&block->b_count);
67         } else {
68                 list_del_init(&block->b_list);
69         }
70
71         pos = &nlm_blocked;
72         if (when != NLM_NEVER) {
73                 if ((when += jiffies) == NLM_NEVER)
74                         when ++;
75                 list_for_each(pos, &nlm_blocked) {
76                         b = list_entry(pos, struct nlm_block, b_list);
77                         if (time_after(b->b_when,when) || b->b_when == NLM_NEVER)
78                                 break;
79                 }
80                 /* On normal exit from the loop, pos == &nlm_blocked,
81                  * so we will be adding to the end of the list - good
82                  */
83         }
84
85         list_add_tail(&block->b_list, pos);
86         block->b_when = when;
87 }
88
89 /*
90  * Remove a block from the global list
91  */
92 static inline void
93 nlmsvc_remove_block(struct nlm_block *block)
94 {
95         if (!list_empty(&block->b_list)) {
96                 list_del_init(&block->b_list);
97                 nlmsvc_release_block(block);
98         }
99 }
100
101 /*
102  * Find a block for a given lock
103  */
104 static struct nlm_block *
105 nlmsvc_lookup_block(struct nlm_file *file, struct nlm_lock *lock)
106 {
107         struct nlm_block        *block;
108         struct file_lock        *fl;
109
110         dprintk("lockd: nlmsvc_lookup_block f=%p pd=%d %Ld-%Ld ty=%d\n",
111                                 file, lock->fl.fl_pid,
112                                 (long long)lock->fl.fl_start,
113                                 (long long)lock->fl.fl_end, lock->fl.fl_type);
114         list_for_each_entry(block, &nlm_blocked, b_list) {
115                 fl = &block->b_call->a_args.lock.fl;
116                 dprintk("lockd: check f=%p pd=%d %Ld-%Ld ty=%d cookie=%s\n",
117                                 block->b_file, fl->fl_pid,
118                                 (long long)fl->fl_start,
119                                 (long long)fl->fl_end, fl->fl_type,
120                                 nlmdbg_cookie2a(&block->b_call->a_args.cookie));
121                 if (block->b_file == file && nlm_compare_locks(fl, &lock->fl)) {
122                         kref_get(&block->b_count);
123                         return block;
124                 }
125         }
126
127         return NULL;
128 }
129
130 static inline int nlm_cookie_match(struct nlm_cookie *a, struct nlm_cookie *b)
131 {
132         if(a->len != b->len)
133                 return 0;
134         if(memcmp(a->data,b->data,a->len))
135                 return 0;
136         return 1;
137 }
138
139 /*
140  * Find a block with a given NLM cookie.
141  */
142 static inline struct nlm_block *
143 nlmsvc_find_block(struct nlm_cookie *cookie)
144 {
145         struct nlm_block *block;
146
147         list_for_each_entry(block, &nlm_blocked, b_list) {
148                 if (nlm_cookie_match(&block->b_call->a_args.cookie,cookie))
149                         goto found;
150         }
151
152         return NULL;
153
154 found:
155         dprintk("nlmsvc_find_block(%s): block=%p\n", nlmdbg_cookie2a(cookie), block);
156         kref_get(&block->b_count);
157         return block;
158 }
159
160 /*
161  * Create a block and initialize it.
162  *
163  * Note: we explicitly set the cookie of the grant reply to that of
164  * the blocked lock request. The spec explicitly mentions that the client
165  * should _not_ rely on the callback containing the same cookie as the
166  * request, but (as I found out later) that's because some implementations
167  * do just this. Never mind the standards comittees, they support our
168  * logging industries.
169  *
170  * 10 years later: I hope we can safely ignore these old and broken
171  * clients by now. Let's fix this so we can uniquely identify an incoming
172  * GRANTED_RES message by cookie, without having to rely on the client's IP
173  * address. --okir
174  */
175 static struct nlm_block *
176 nlmsvc_create_block(struct svc_rqst *rqstp, struct nlm_host *host,
177                     struct nlm_file *file, struct nlm_lock *lock,
178                     struct nlm_cookie *cookie)
179 {
180         struct nlm_block        *block;
181         struct nlm_rqst         *call = NULL;
182
183         call = nlm_alloc_call(host);
184         if (call == NULL)
185                 return NULL;
186
187         /* Allocate memory for block, and initialize arguments */
188         block = kzalloc(sizeof(*block), GFP_KERNEL);
189         if (block == NULL)
190                 goto failed;
191         kref_init(&block->b_count);
192         INIT_LIST_HEAD(&block->b_list);
193         INIT_LIST_HEAD(&block->b_flist);
194
195         if (!nlmsvc_setgrantargs(call, lock))
196                 goto failed_free;
197
198         /* Set notifier function for VFS, and init args */
199         call->a_args.lock.fl.fl_flags |= FL_SLEEP;
200         call->a_args.lock.fl.fl_lmops = &nlmsvc_lock_operations;
201         nlmclnt_next_cookie(&call->a_args.cookie);
202
203         dprintk("lockd: created block %p...\n", block);
204
205         /* Create and initialize the block */
206         block->b_daemon = rqstp->rq_server;
207         block->b_host   = host;
208         block->b_file   = file;
209         block->b_fl = NULL;
210         file->f_count++;
211
212         /* Add to file's list of blocks */
213         list_add(&block->b_flist, &file->f_blocks);
214
215         /* Set up RPC arguments for callback */
216         block->b_call = call;
217         call->a_flags   = RPC_TASK_ASYNC;
218         call->a_block = block;
219
220         return block;
221
222 failed_free:
223         kfree(block);
224 failed:
225         nlm_release_call(call);
226         return NULL;
227 }
228
229 /*
230  * Delete a block.
231  * It is the caller's responsibility to check whether the file
232  * can be closed hereafter.
233  */
234 static int nlmsvc_unlink_block(struct nlm_block *block)
235 {
236         int status;
237         dprintk("lockd: unlinking block %p...\n", block);
238
239         /* Remove block from list */
240         status = posix_unblock_lock(block->b_file->f_file, &block->b_call->a_args.lock.fl);
241         nlmsvc_remove_block(block);
242         return status;
243 }
244
245 static void nlmsvc_free_block(struct kref *kref)
246 {
247         struct nlm_block *block = container_of(kref, struct nlm_block, b_count);
248         struct nlm_file         *file = block->b_file;
249
250         dprintk("lockd: freeing block %p...\n", block);
251
252         /* Remove block from file's list of blocks */
253         mutex_lock(&file->f_mutex);
254         list_del_init(&block->b_flist);
255         mutex_unlock(&file->f_mutex);
256
257         nlmsvc_freegrantargs(block->b_call);
258         nlm_release_call(block->b_call);
259         nlm_release_file(block->b_file);
260         kfree(block->b_fl);
261         kfree(block);
262 }
263
264 static void nlmsvc_release_block(struct nlm_block *block)
265 {
266         if (block != NULL)
267                 kref_put(&block->b_count, nlmsvc_free_block);
268 }
269
270 /*
271  * Loop over all blocks and delete blocks held by
272  * a matching host.
273  */
274 void nlmsvc_traverse_blocks(struct nlm_host *host,
275                         struct nlm_file *file,
276                         nlm_host_match_fn_t match)
277 {
278         struct nlm_block *block, *next;
279
280 restart:
281         mutex_lock(&file->f_mutex);
282         list_for_each_entry_safe(block, next, &file->f_blocks, b_flist) {
283                 if (!match(block->b_host, host))
284                         continue;
285                 /* Do not destroy blocks that are not on
286                  * the global retry list - why? */
287                 if (list_empty(&block->b_list))
288                         continue;
289                 kref_get(&block->b_count);
290                 mutex_unlock(&file->f_mutex);
291                 nlmsvc_unlink_block(block);
292                 nlmsvc_release_block(block);
293                 goto restart;
294         }
295         mutex_unlock(&file->f_mutex);
296 }
297
298 /*
299  * Initialize arguments for GRANTED call. The nlm_rqst structure
300  * has been cleared already.
301  */
302 static int nlmsvc_setgrantargs(struct nlm_rqst *call, struct nlm_lock *lock)
303 {
304         locks_copy_lock(&call->a_args.lock.fl, &lock->fl);
305         memcpy(&call->a_args.lock.fh, &lock->fh, sizeof(call->a_args.lock.fh));
306         call->a_args.lock.caller = utsname()->nodename;
307         call->a_args.lock.oh.len = lock->oh.len;
308
309         /* set default data area */
310         call->a_args.lock.oh.data = call->a_owner;
311         call->a_args.lock.svid = lock->fl.fl_pid;
312
313         if (lock->oh.len > NLMCLNT_OHSIZE) {
314                 void *data = kmalloc(lock->oh.len, GFP_KERNEL);
315                 if (!data)
316                         return 0;
317                 call->a_args.lock.oh.data = (u8 *) data;
318         }
319
320         memcpy(call->a_args.lock.oh.data, lock->oh.data, lock->oh.len);
321         return 1;
322 }
323
324 static void nlmsvc_freegrantargs(struct nlm_rqst *call)
325 {
326         if (call->a_args.lock.oh.data != call->a_owner)
327                 kfree(call->a_args.lock.oh.data);
328 }
329
330 /*
331  * Deferred lock request handling for non-blocking lock
332  */
333 static __be32
334 nlmsvc_defer_lock_rqst(struct svc_rqst *rqstp, struct nlm_block *block)
335 {
336         __be32 status = nlm_lck_denied_nolocks;
337
338         block->b_flags |= B_QUEUED;
339
340         nlmsvc_insert_block(block, NLM_TIMEOUT);
341
342         block->b_cache_req = &rqstp->rq_chandle;
343         if (rqstp->rq_chandle.defer) {
344                 block->b_deferred_req =
345                         rqstp->rq_chandle.defer(block->b_cache_req);
346                 if (block->b_deferred_req != NULL)
347                         status = nlm_drop_reply;
348         }
349         dprintk("lockd: nlmsvc_defer_lock_rqst block %p flags %d status %d\n",
350                 block, block->b_flags, ntohl(status));
351
352         return status;
353 }
354
355 /*
356  * Attempt to establish a lock, and if it can't be granted, block it
357  * if required.
358  */
359 __be32
360 nlmsvc_lock(struct svc_rqst *rqstp, struct nlm_file *file,
361                         struct nlm_lock *lock, int wait, struct nlm_cookie *cookie)
362 {
363         struct nlm_block        *block = NULL;
364         struct nlm_host         *host;
365         int                     error;
366         __be32                  ret;
367
368         dprintk("lockd: nlmsvc_lock(%s/%ld, ty=%d, pi=%d, %Ld-%Ld, bl=%d)\n",
369                                 file->f_file->f_path.dentry->d_inode->i_sb->s_id,
370                                 file->f_file->f_path.dentry->d_inode->i_ino,
371                                 lock->fl.fl_type, lock->fl.fl_pid,
372                                 (long long)lock->fl.fl_start,
373                                 (long long)lock->fl.fl_end,
374                                 wait);
375
376         /* Create host handle for callback */
377         host = nlmsvc_lookup_host(rqstp, lock->caller, lock->len);
378         if (host == NULL)
379                 return nlm_lck_denied_nolocks;
380
381         /* Lock file against concurrent access */
382         mutex_lock(&file->f_mutex);
383         /* Get existing block (in case client is busy-waiting)
384          * or create new block
385          */
386         block = nlmsvc_lookup_block(file, lock);
387         if (block == NULL) {
388                 block = nlmsvc_create_block(rqstp, nlm_get_host(host), file,
389                                 lock, cookie);
390                 ret = nlm_lck_denied_nolocks;
391                 if (block == NULL)
392                         goto out;
393                 lock = &block->b_call->a_args.lock;
394         } else
395                 lock->fl.fl_flags &= ~FL_SLEEP;
396
397         if (block->b_flags & B_QUEUED) {
398                 dprintk("lockd: nlmsvc_lock deferred block %p flags %d\n",
399                                                         block, block->b_flags);
400                 if (block->b_granted) {
401                         nlmsvc_unlink_block(block);
402                         ret = nlm_granted;
403                         goto out;
404                 }
405                 if (block->b_flags & B_TIMED_OUT) {
406                         nlmsvc_unlink_block(block);
407                         ret = nlm_lck_denied;
408                         goto out;
409                 }
410                 ret = nlm_drop_reply;
411                 goto out;
412         }
413
414         if (!wait)
415                 lock->fl.fl_flags &= ~FL_SLEEP;
416         error = vfs_lock_file(file->f_file, F_SETLK, &lock->fl, NULL);
417         lock->fl.fl_flags &= ~FL_SLEEP;
418
419         dprintk("lockd: vfs_lock_file returned %d\n", error);
420         switch(error) {
421                 case 0:
422                         ret = nlm_granted;
423                         goto out;
424                 case -EAGAIN:
425                         ret = nlm_lck_denied;
426                         break;
427                 case -EINPROGRESS:
428                         if (wait)
429                                 break;
430                         /* Filesystem lock operation is in progress
431                            Add it to the queue waiting for callback */
432                         ret = nlmsvc_defer_lock_rqst(rqstp, block);
433                         goto out;
434                 case -EDEADLK:
435                         ret = nlm_deadlock;
436                         goto out;
437                 default:                        /* includes ENOLCK */
438                         ret = nlm_lck_denied_nolocks;
439                         goto out;
440         }
441
442         ret = nlm_lck_denied;
443         if (!wait)
444                 goto out;
445
446         ret = nlm_lck_blocked;
447
448         /* Append to list of blocked */
449         nlmsvc_insert_block(block, NLM_NEVER);
450 out:
451         mutex_unlock(&file->f_mutex);
452         nlmsvc_release_block(block);
453         nlm_release_host(host);
454         dprintk("lockd: nlmsvc_lock returned %u\n", ret);
455         return ret;
456 }
457
458 /*
459  * Test for presence of a conflicting lock.
460  */
461 __be32
462 nlmsvc_testlock(struct svc_rqst *rqstp, struct nlm_file *file,
463                 struct nlm_lock *lock, struct nlm_lock *conflock,
464                 struct nlm_cookie *cookie)
465 {
466         struct nlm_block        *block = NULL;
467         int                     error;
468         __be32                  ret;
469
470         dprintk("lockd: nlmsvc_testlock(%s/%ld, ty=%d, %Ld-%Ld)\n",
471                                 file->f_file->f_path.dentry->d_inode->i_sb->s_id,
472                                 file->f_file->f_path.dentry->d_inode->i_ino,
473                                 lock->fl.fl_type,
474                                 (long long)lock->fl.fl_start,
475                                 (long long)lock->fl.fl_end);
476
477         /* Get existing block (in case client is busy-waiting) */
478         block = nlmsvc_lookup_block(file, lock);
479
480         if (block == NULL) {
481                 struct file_lock *conf = kzalloc(sizeof(*conf), GFP_KERNEL);
482                 struct nlm_host *host;
483
484                 if (conf == NULL)
485                         return nlm_granted;
486                 /* Create host handle for callback */
487                 host = nlmsvc_lookup_host(rqstp, lock->caller, lock->len);
488                 if (host == NULL) {
489                         kfree(conf);
490                         return nlm_lck_denied_nolocks;
491                 }
492                 block = nlmsvc_create_block(rqstp, host, file, lock, cookie);
493                 if (block == NULL) {
494                         kfree(conf);
495                         return nlm_granted;
496                 }
497                 block->b_fl = conf;
498         }
499         if (block->b_flags & B_QUEUED) {
500                 dprintk("lockd: nlmsvc_testlock deferred block %p flags %d fl %p\n",
501                         block, block->b_flags, block->b_fl);
502                 if (block->b_flags & B_TIMED_OUT) {
503                         nlmsvc_unlink_block(block);
504                         ret = nlm_lck_denied;
505                         goto out;
506                 }
507                 if (block->b_flags & B_GOT_CALLBACK) {
508                         nlmsvc_unlink_block(block);
509                         if (block->b_fl != NULL
510                                         && block->b_fl->fl_type != F_UNLCK) {
511                                 lock->fl = *block->b_fl;
512                                 goto conf_lock;
513                         } else {
514                                 ret = nlm_granted;
515                                 goto out;
516                         }
517                 }
518                 ret = nlm_drop_reply;
519                 goto out;
520         }
521
522         error = vfs_test_lock(file->f_file, &lock->fl);
523         if (error == -EINPROGRESS) {
524                 ret = nlmsvc_defer_lock_rqst(rqstp, block);
525                 goto out;
526         }
527         if (error) {
528                 ret = nlm_lck_denied_nolocks;
529                 goto out;
530         }
531         if (lock->fl.fl_type == F_UNLCK) {
532                 ret = nlm_granted;
533                 goto out;
534         }
535
536 conf_lock:
537         dprintk("lockd: conflicting lock(ty=%d, %Ld-%Ld)\n",
538                 lock->fl.fl_type, (long long)lock->fl.fl_start,
539                 (long long)lock->fl.fl_end);
540         conflock->caller = "somehost";  /* FIXME */
541         conflock->len = strlen(conflock->caller);
542         conflock->oh.len = 0;           /* don't return OH info */
543         conflock->svid = lock->fl.fl_pid;
544         conflock->fl.fl_type = lock->fl.fl_type;
545         conflock->fl.fl_start = lock->fl.fl_start;
546         conflock->fl.fl_end = lock->fl.fl_end;
547         ret = nlm_lck_denied;
548 out:
549         if (block)
550                 nlmsvc_release_block(block);
551         return ret;
552 }
553
554 /*
555  * Remove a lock.
556  * This implies a CANCEL call: We send a GRANT_MSG, the client replies
557  * with a GRANT_RES call which gets lost, and calls UNLOCK immediately
558  * afterwards. In this case the block will still be there, and hence
559  * must be removed.
560  */
561 __be32
562 nlmsvc_unlock(struct nlm_file *file, struct nlm_lock *lock)
563 {
564         int     error;
565
566         dprintk("lockd: nlmsvc_unlock(%s/%ld, pi=%d, %Ld-%Ld)\n",
567                                 file->f_file->f_path.dentry->d_inode->i_sb->s_id,
568                                 file->f_file->f_path.dentry->d_inode->i_ino,
569                                 lock->fl.fl_pid,
570                                 (long long)lock->fl.fl_start,
571                                 (long long)lock->fl.fl_end);
572
573         /* First, cancel any lock that might be there */
574         nlmsvc_cancel_blocked(file, lock);
575
576         lock->fl.fl_type = F_UNLCK;
577         error = vfs_lock_file(file->f_file, F_SETLK, &lock->fl, NULL);
578
579         return (error < 0)? nlm_lck_denied_nolocks : nlm_granted;
580 }
581
582 /*
583  * Cancel a previously blocked request.
584  *
585  * A cancel request always overrides any grant that may currently
586  * be in progress.
587  * The calling procedure must check whether the file can be closed.
588  */
589 __be32
590 nlmsvc_cancel_blocked(struct nlm_file *file, struct nlm_lock *lock)
591 {
592         struct nlm_block        *block;
593         int status = 0;
594
595         dprintk("lockd: nlmsvc_cancel(%s/%ld, pi=%d, %Ld-%Ld)\n",
596                                 file->f_file->f_path.dentry->d_inode->i_sb->s_id,
597                                 file->f_file->f_path.dentry->d_inode->i_ino,
598                                 lock->fl.fl_pid,
599                                 (long long)lock->fl.fl_start,
600                                 (long long)lock->fl.fl_end);
601
602         mutex_lock(&file->f_mutex);
603         block = nlmsvc_lookup_block(file, lock);
604         mutex_unlock(&file->f_mutex);
605         if (block != NULL) {
606                 vfs_cancel_lock(block->b_file->f_file,
607                                 &block->b_call->a_args.lock.fl);
608                 status = nlmsvc_unlink_block(block);
609                 nlmsvc_release_block(block);
610         }
611         return status ? nlm_lck_denied : nlm_granted;
612 }
613
614 /*
615  * This is a callback from the filesystem for VFS file lock requests.
616  * It will be used if fl_grant is defined and the filesystem can not
617  * respond to the request immediately.
618  * For GETLK request it will copy the reply to the nlm_block.
619  * For SETLK or SETLKW request it will get the local posix lock.
620  * In all cases it will move the block to the head of nlm_blocked q where
621  * nlmsvc_retry_blocked() can send back a reply for SETLKW or revisit the
622  * deferred rpc for GETLK and SETLK.
623  */
624 static void
625 nlmsvc_update_deferred_block(struct nlm_block *block, struct file_lock *conf,
626                              int result)
627 {
628         block->b_flags |= B_GOT_CALLBACK;
629         if (result == 0)
630                 block->b_granted = 1;
631         else
632                 block->b_flags |= B_TIMED_OUT;
633         if (conf) {
634                 if (block->b_fl)
635                         __locks_copy_lock(block->b_fl, conf);
636         }
637 }
638
639 static int nlmsvc_grant_deferred(struct file_lock *fl, struct file_lock *conf,
640                                         int result)
641 {
642         struct nlm_block *block;
643         int rc = -ENOENT;
644
645         lock_kernel();
646         list_for_each_entry(block, &nlm_blocked, b_list) {
647                 if (nlm_compare_locks(&block->b_call->a_args.lock.fl, fl)) {
648                         dprintk("lockd: nlmsvc_notify_blocked block %p flags %d\n",
649                                                         block, block->b_flags);
650                         if (block->b_flags & B_QUEUED) {
651                                 if (block->b_flags & B_TIMED_OUT) {
652                                         rc = -ENOLCK;
653                                         break;
654                                 }
655                                 nlmsvc_update_deferred_block(block, conf, result);
656                         } else if (result == 0)
657                                 block->b_granted = 1;
658
659                         nlmsvc_insert_block(block, 0);
660                         svc_wake_up(block->b_daemon);
661                         rc = 0;
662                         break;
663                 }
664         }
665         unlock_kernel();
666         if (rc == -ENOENT)
667                 printk(KERN_WARNING "lockd: grant for unknown block\n");
668         return rc;
669 }
670
671 /*
672  * Unblock a blocked lock request. This is a callback invoked from the
673  * VFS layer when a lock on which we blocked is removed.
674  *
675  * This function doesn't grant the blocked lock instantly, but rather moves
676  * the block to the head of nlm_blocked where it can be picked up by lockd.
677  */
678 static void
679 nlmsvc_notify_blocked(struct file_lock *fl)
680 {
681         struct nlm_block        *block;
682
683         dprintk("lockd: VFS unblock notification for block %p\n", fl);
684         list_for_each_entry(block, &nlm_blocked, b_list) {
685                 if (nlm_compare_locks(&block->b_call->a_args.lock.fl, fl)) {
686                         nlmsvc_insert_block(block, 0);
687                         svc_wake_up(block->b_daemon);
688                         return;
689                 }
690         }
691
692         printk(KERN_WARNING "lockd: notification for unknown block!\n");
693 }
694
695 static int nlmsvc_same_owner(struct file_lock *fl1, struct file_lock *fl2)
696 {
697         return fl1->fl_owner == fl2->fl_owner && fl1->fl_pid == fl2->fl_pid;
698 }
699
700 struct lock_manager_operations nlmsvc_lock_operations = {
701         .fl_compare_owner = nlmsvc_same_owner,
702         .fl_notify = nlmsvc_notify_blocked,
703         .fl_grant = nlmsvc_grant_deferred,
704 };
705
706 /*
707  * Try to claim a lock that was previously blocked.
708  *
709  * Note that we use both the RPC_GRANTED_MSG call _and_ an async
710  * RPC thread when notifying the client. This seems like overkill...
711  * Here's why:
712  *  -   we don't want to use a synchronous RPC thread, otherwise
713  *      we might find ourselves hanging on a dead portmapper.
714  *  -   Some lockd implementations (e.g. HP) don't react to
715  *      RPC_GRANTED calls; they seem to insist on RPC_GRANTED_MSG calls.
716  */
717 static void
718 nlmsvc_grant_blocked(struct nlm_block *block)
719 {
720         struct nlm_file         *file = block->b_file;
721         struct nlm_lock         *lock = &block->b_call->a_args.lock;
722         int                     error;
723
724         dprintk("lockd: grant blocked lock %p\n", block);
725
726         kref_get(&block->b_count);
727
728         /* Unlink block request from list */
729         nlmsvc_unlink_block(block);
730
731         /* If b_granted is true this means we've been here before.
732          * Just retry the grant callback, possibly refreshing the RPC
733          * binding */
734         if (block->b_granted) {
735                 nlm_rebind_host(block->b_host);
736                 goto callback;
737         }
738
739         /* Try the lock operation again */
740         lock->fl.fl_flags |= FL_SLEEP;
741         error = vfs_lock_file(file->f_file, F_SETLK, &lock->fl, NULL);
742         lock->fl.fl_flags &= ~FL_SLEEP;
743
744         switch (error) {
745         case 0:
746                 break;
747         case -EAGAIN:
748         case -EINPROGRESS:
749                 dprintk("lockd: lock still blocked error %d\n", error);
750                 nlmsvc_insert_block(block, NLM_NEVER);
751                 nlmsvc_release_block(block);
752                 return;
753         default:
754                 printk(KERN_WARNING "lockd: unexpected error %d in %s!\n",
755                                 -error, __func__);
756                 nlmsvc_insert_block(block, 10 * HZ);
757                 nlmsvc_release_block(block);
758                 return;
759         }
760
761 callback:
762         /* Lock was granted by VFS. */
763         dprintk("lockd: GRANTing blocked lock.\n");
764         block->b_granted = 1;
765
766         /* keep block on the list, but don't reattempt until the RPC
767          * completes or the submission fails
768          */
769         nlmsvc_insert_block(block, NLM_NEVER);
770
771         /* Call the client -- use a soft RPC task since nlmsvc_retry_blocked
772          * will queue up a new one if this one times out
773          */
774         error = nlm_async_call(block->b_call, NLMPROC_GRANTED_MSG,
775                                 &nlmsvc_grant_ops);
776
777         /* RPC submission failed, wait a bit and retry */
778         if (error < 0)
779                 nlmsvc_insert_block(block, 10 * HZ);
780 }
781
782 /*
783  * This is the callback from the RPC layer when the NLM_GRANTED_MSG
784  * RPC call has succeeded or timed out.
785  * Like all RPC callbacks, it is invoked by the rpciod process, so it
786  * better not sleep. Therefore, we put the blocked lock on the nlm_blocked
787  * chain once more in order to have it removed by lockd itself (which can
788  * then sleep on the file semaphore without disrupting e.g. the nfs client).
789  */
790 static void nlmsvc_grant_callback(struct rpc_task *task, void *data)
791 {
792         struct nlm_rqst         *call = data;
793         struct nlm_block        *block = call->a_block;
794         unsigned long           timeout;
795
796         dprintk("lockd: GRANT_MSG RPC callback\n");
797
798         /* if the block is not on a list at this point then it has
799          * been invalidated. Don't try to requeue it.
800          *
801          * FIXME: it's possible that the block is removed from the list
802          * after this check but before the nlmsvc_insert_block. In that
803          * case it will be added back. Perhaps we need better locking
804          * for nlm_blocked?
805          */
806         if (list_empty(&block->b_list))
807                 return;
808
809         /* Technically, we should down the file semaphore here. Since we
810          * move the block towards the head of the queue only, no harm
811          * can be done, though. */
812         if (task->tk_status < 0) {
813                 /* RPC error: Re-insert for retransmission */
814                 timeout = 10 * HZ;
815         } else {
816                 /* Call was successful, now wait for client callback */
817                 timeout = 60 * HZ;
818         }
819         nlmsvc_insert_block(block, timeout);
820         svc_wake_up(block->b_daemon);
821 }
822
823 static void nlmsvc_grant_release(void *data)
824 {
825         struct nlm_rqst         *call = data;
826
827         nlmsvc_release_block(call->a_block);
828 }
829
830 static const struct rpc_call_ops nlmsvc_grant_ops = {
831         .rpc_call_done = nlmsvc_grant_callback,
832         .rpc_release = nlmsvc_grant_release,
833 };
834
835 /*
836  * We received a GRANT_RES callback. Try to find the corresponding
837  * block.
838  */
839 void
840 nlmsvc_grant_reply(struct nlm_cookie *cookie, __be32 status)
841 {
842         struct nlm_block        *block;
843
844         dprintk("grant_reply: looking for cookie %x, s=%d \n",
845                 *(unsigned int *)(cookie->data), status);
846         if (!(block = nlmsvc_find_block(cookie)))
847                 return;
848
849         if (block) {
850                 if (status == nlm_lck_denied_grace_period) {
851                         /* Try again in a couple of seconds */
852                         nlmsvc_insert_block(block, 10 * HZ);
853                 } else {
854                         /* Lock is now held by client, or has been rejected.
855                          * In both cases, the block should be removed. */
856                         nlmsvc_unlink_block(block);
857                 }
858         }
859         nlmsvc_release_block(block);
860 }
861
862 /* Helper function to handle retry of a deferred block.
863  * If it is a blocking lock, call grant_blocked.
864  * For a non-blocking lock or test lock, revisit the request.
865  */
866 static void
867 retry_deferred_block(struct nlm_block *block)
868 {
869         if (!(block->b_flags & B_GOT_CALLBACK))
870                 block->b_flags |= B_TIMED_OUT;
871         nlmsvc_insert_block(block, NLM_TIMEOUT);
872         dprintk("revisit block %p flags %d\n",  block, block->b_flags);
873         if (block->b_deferred_req) {
874                 block->b_deferred_req->revisit(block->b_deferred_req, 0);
875                 block->b_deferred_req = NULL;
876         }
877 }
878
879 /*
880  * Retry all blocked locks that have been notified. This is where lockd
881  * picks up locks that can be granted, or grant notifications that must
882  * be retransmitted.
883  */
884 unsigned long
885 nlmsvc_retry_blocked(void)
886 {
887         unsigned long   timeout = MAX_SCHEDULE_TIMEOUT;
888         struct nlm_block *block;
889
890         while (!list_empty(&nlm_blocked) && !kthread_should_stop()) {
891                 block = list_entry(nlm_blocked.next, struct nlm_block, b_list);
892
893                 if (block->b_when == NLM_NEVER)
894                         break;
895                 if (time_after(block->b_when,jiffies)) {
896                         timeout = block->b_when - jiffies;
897                         break;
898                 }
899
900                 dprintk("nlmsvc_retry_blocked(%p, when=%ld)\n",
901                         block, block->b_when);
902                 if (block->b_flags & B_QUEUED) {
903                         dprintk("nlmsvc_retry_blocked delete block (%p, granted=%d, flags=%d)\n",
904                                 block, block->b_granted, block->b_flags);
905                         retry_deferred_block(block);
906                 } else
907                         nlmsvc_grant_blocked(block);
908         }
909
910         return timeout;
911 }