NLM: Convert lockd to use kthreads
[linux-2.6.git] / 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. If the lock was cancelled or the grant callback
231  * failed, unlock is set to 1.
232  * It is the caller's responsibility to check whether the file
233  * can be closed hereafter.
234  */
235 static int nlmsvc_unlink_block(struct nlm_block *block)
236 {
237         int status;
238         dprintk("lockd: unlinking block %p...\n", block);
239
240         /* Remove block from list */
241         status = posix_unblock_lock(block->b_file->f_file, &block->b_call->a_args.lock.fl);
242         nlmsvc_remove_block(block);
243         return status;
244 }
245
246 static void nlmsvc_free_block(struct kref *kref)
247 {
248         struct nlm_block *block = container_of(kref, struct nlm_block, b_count);
249         struct nlm_file         *file = block->b_file;
250
251         dprintk("lockd: freeing block %p...\n", block);
252
253         /* Remove block from file's list of blocks */
254         mutex_lock(&file->f_mutex);
255         list_del_init(&block->b_flist);
256         mutex_unlock(&file->f_mutex);
257
258         nlmsvc_freegrantargs(block->b_call);
259         nlm_release_call(block->b_call);
260         nlm_release_file(block->b_file);
261         kfree(block->b_fl);
262         kfree(block);
263 }
264
265 static void nlmsvc_release_block(struct nlm_block *block)
266 {
267         if (block != NULL)
268                 kref_put(&block->b_count, nlmsvc_free_block);
269 }
270
271 /*
272  * Loop over all blocks and delete blocks held by
273  * a matching host.
274  */
275 void nlmsvc_traverse_blocks(struct nlm_host *host,
276                         struct nlm_file *file,
277                         nlm_host_match_fn_t match)
278 {
279         struct nlm_block *block, *next;
280
281 restart:
282         mutex_lock(&file->f_mutex);
283         list_for_each_entry_safe(block, next, &file->f_blocks, b_flist) {
284                 if (!match(block->b_host, host))
285                         continue;
286                 /* Do not destroy blocks that are not on
287                  * the global retry list - why? */
288                 if (list_empty(&block->b_list))
289                         continue;
290                 kref_get(&block->b_count);
291                 mutex_unlock(&file->f_mutex);
292                 nlmsvc_unlink_block(block);
293                 nlmsvc_release_block(block);
294                 goto restart;
295         }
296         mutex_unlock(&file->f_mutex);
297 }
298
299 /*
300  * Initialize arguments for GRANTED call. The nlm_rqst structure
301  * has been cleared already.
302  */
303 static int nlmsvc_setgrantargs(struct nlm_rqst *call, struct nlm_lock *lock)
304 {
305         locks_copy_lock(&call->a_args.lock.fl, &lock->fl);
306         memcpy(&call->a_args.lock.fh, &lock->fh, sizeof(call->a_args.lock.fh));
307         call->a_args.lock.caller = utsname()->nodename;
308         call->a_args.lock.oh.len = lock->oh.len;
309
310         /* set default data area */
311         call->a_args.lock.oh.data = call->a_owner;
312         call->a_args.lock.svid = lock->fl.fl_pid;
313
314         if (lock->oh.len > NLMCLNT_OHSIZE) {
315                 void *data = kmalloc(lock->oh.len, GFP_KERNEL);
316                 if (!data)
317                         return 0;
318                 call->a_args.lock.oh.data = (u8 *) data;
319         }
320
321         memcpy(call->a_args.lock.oh.data, lock->oh.data, lock->oh.len);
322         return 1;
323 }
324
325 static void nlmsvc_freegrantargs(struct nlm_rqst *call)
326 {
327         if (call->a_args.lock.oh.data != call->a_owner)
328                 kfree(call->a_args.lock.oh.data);
329 }
330
331 /*
332  * Deferred lock request handling for non-blocking lock
333  */
334 static __be32
335 nlmsvc_defer_lock_rqst(struct svc_rqst *rqstp, struct nlm_block *block)
336 {
337         __be32 status = nlm_lck_denied_nolocks;
338
339         block->b_flags |= B_QUEUED;
340
341         nlmsvc_insert_block(block, NLM_TIMEOUT);
342
343         block->b_cache_req = &rqstp->rq_chandle;
344         if (rqstp->rq_chandle.defer) {
345                 block->b_deferred_req =
346                         rqstp->rq_chandle.defer(block->b_cache_req);
347                 if (block->b_deferred_req != NULL)
348                         status = nlm_drop_reply;
349         }
350         dprintk("lockd: nlmsvc_defer_lock_rqst block %p flags %d status %d\n",
351                 block, block->b_flags, ntohl(status));
352
353         return status;
354 }
355
356 /*
357  * Attempt to establish a lock, and if it can't be granted, block it
358  * if required.
359  */
360 __be32
361 nlmsvc_lock(struct svc_rqst *rqstp, struct nlm_file *file,
362                         struct nlm_lock *lock, int wait, struct nlm_cookie *cookie)
363 {
364         struct nlm_block        *block = NULL;
365         struct nlm_host         *host;
366         int                     error;
367         __be32                  ret;
368
369         dprintk("lockd: nlmsvc_lock(%s/%ld, ty=%d, pi=%d, %Ld-%Ld, bl=%d)\n",
370                                 file->f_file->f_path.dentry->d_inode->i_sb->s_id,
371                                 file->f_file->f_path.dentry->d_inode->i_ino,
372                                 lock->fl.fl_type, lock->fl.fl_pid,
373                                 (long long)lock->fl.fl_start,
374                                 (long long)lock->fl.fl_end,
375                                 wait);
376
377         /* Create host handle for callback */
378         host = nlmsvc_lookup_host(rqstp, lock->caller, lock->len);
379         if (host == NULL)
380                 return nlm_lck_denied_nolocks;
381
382         /* Lock file against concurrent access */
383         mutex_lock(&file->f_mutex);
384         /* Get existing block (in case client is busy-waiting)
385          * or create new block
386          */
387         block = nlmsvc_lookup_block(file, lock);
388         if (block == NULL) {
389                 block = nlmsvc_create_block(rqstp, nlm_get_host(host), file,
390                                 lock, cookie);
391                 ret = nlm_lck_denied_nolocks;
392                 if (block == NULL)
393                         goto out;
394                 lock = &block->b_call->a_args.lock;
395         } else
396                 lock->fl.fl_flags &= ~FL_SLEEP;
397
398         if (block->b_flags & B_QUEUED) {
399                 dprintk("lockd: nlmsvc_lock deferred block %p flags %d\n",
400                                                         block, block->b_flags);
401                 if (block->b_granted) {
402                         nlmsvc_unlink_block(block);
403                         ret = nlm_granted;
404                         goto out;
405                 }
406                 if (block->b_flags & B_TIMED_OUT) {
407                         nlmsvc_unlink_block(block);
408                         ret = nlm_lck_denied;
409                         goto out;
410                 }
411                 ret = nlm_drop_reply;
412                 goto out;
413         }
414
415         if (!wait)
416                 lock->fl.fl_flags &= ~FL_SLEEP;
417         error = vfs_lock_file(file->f_file, F_SETLK, &lock->fl, NULL);
418         lock->fl.fl_flags &= ~FL_SLEEP;
419
420         dprintk("lockd: vfs_lock_file returned %d\n", error);
421         switch(error) {
422                 case 0:
423                         ret = nlm_granted;
424                         goto out;
425                 case -EAGAIN:
426                         ret = nlm_lck_denied;
427                         break;
428                 case -EINPROGRESS:
429                         if (wait)
430                                 break;
431                         /* Filesystem lock operation is in progress
432                            Add it to the queue waiting for callback */
433                         ret = nlmsvc_defer_lock_rqst(rqstp, block);
434                         goto out;
435                 case -EDEADLK:
436                         ret = nlm_deadlock;
437                         goto out;
438                 default:                        /* includes ENOLCK */
439                         ret = nlm_lck_denied_nolocks;
440                         goto out;
441         }
442
443         ret = nlm_lck_denied;
444         if (!wait)
445                 goto out;
446
447         ret = nlm_lck_blocked;
448
449         /* Append to list of blocked */
450         nlmsvc_insert_block(block, NLM_NEVER);
451 out:
452         mutex_unlock(&file->f_mutex);
453         nlmsvc_release_block(block);
454         nlm_release_host(host);
455         dprintk("lockd: nlmsvc_lock returned %u\n", ret);
456         return ret;
457 }
458
459 /*
460  * Test for presence of a conflicting lock.
461  */
462 __be32
463 nlmsvc_testlock(struct svc_rqst *rqstp, struct nlm_file *file,
464                 struct nlm_lock *lock, struct nlm_lock *conflock,
465                 struct nlm_cookie *cookie)
466 {
467         struct nlm_block        *block = NULL;
468         int                     error;
469         __be32                  ret;
470
471         dprintk("lockd: nlmsvc_testlock(%s/%ld, ty=%d, %Ld-%Ld)\n",
472                                 file->f_file->f_path.dentry->d_inode->i_sb->s_id,
473                                 file->f_file->f_path.dentry->d_inode->i_ino,
474                                 lock->fl.fl_type,
475                                 (long long)lock->fl.fl_start,
476                                 (long long)lock->fl.fl_end);
477
478         /* Get existing block (in case client is busy-waiting) */
479         block = nlmsvc_lookup_block(file, lock);
480
481         if (block == NULL) {
482                 struct file_lock *conf = kzalloc(sizeof(*conf), GFP_KERNEL);
483                 struct nlm_host *host;
484
485                 if (conf == NULL)
486                         return nlm_granted;
487                 /* Create host handle for callback */
488                 host = nlmsvc_lookup_host(rqstp, lock->caller, lock->len);
489                 if (host == NULL) {
490                         kfree(conf);
491                         return nlm_lck_denied_nolocks;
492                 }
493                 block = nlmsvc_create_block(rqstp, host, file, lock, cookie);
494                 if (block == NULL) {
495                         kfree(conf);
496                         return nlm_granted;
497                 }
498                 block->b_fl = conf;
499         }
500         if (block->b_flags & B_QUEUED) {
501                 dprintk("lockd: nlmsvc_testlock deferred block %p flags %d fl %p\n",
502                         block, block->b_flags, block->b_fl);
503                 if (block->b_flags & B_TIMED_OUT) {
504                         nlmsvc_unlink_block(block);
505                         ret = nlm_lck_denied;
506                         goto out;
507                 }
508                 if (block->b_flags & B_GOT_CALLBACK) {
509                         nlmsvc_unlink_block(block);
510                         if (block->b_fl != NULL
511                                         && block->b_fl->fl_type != F_UNLCK) {
512                                 lock->fl = *block->b_fl;
513                                 goto conf_lock;
514                         } else {
515                                 ret = nlm_granted;
516                                 goto out;
517                         }
518                 }
519                 ret = nlm_drop_reply;
520                 goto out;
521         }
522
523         error = vfs_test_lock(file->f_file, &lock->fl);
524         if (error == -EINPROGRESS) {
525                 ret = nlmsvc_defer_lock_rqst(rqstp, block);
526                 goto out;
527         }
528         if (error) {
529                 ret = nlm_lck_denied_nolocks;
530                 goto out;
531         }
532         if (lock->fl.fl_type == F_UNLCK) {
533                 ret = nlm_granted;
534                 goto out;
535         }
536
537 conf_lock:
538         dprintk("lockd: conflicting lock(ty=%d, %Ld-%Ld)\n",
539                 lock->fl.fl_type, (long long)lock->fl.fl_start,
540                 (long long)lock->fl.fl_end);
541         conflock->caller = "somehost";  /* FIXME */
542         conflock->len = strlen(conflock->caller);
543         conflock->oh.len = 0;           /* don't return OH info */
544         conflock->svid = lock->fl.fl_pid;
545         conflock->fl.fl_type = lock->fl.fl_type;
546         conflock->fl.fl_start = lock->fl.fl_start;
547         conflock->fl.fl_end = lock->fl.fl_end;
548         ret = nlm_lck_denied;
549 out:
550         if (block)
551                 nlmsvc_release_block(block);
552         return ret;
553 }
554
555 /*
556  * Remove a lock.
557  * This implies a CANCEL call: We send a GRANT_MSG, the client replies
558  * with a GRANT_RES call which gets lost, and calls UNLOCK immediately
559  * afterwards. In this case the block will still be there, and hence
560  * must be removed.
561  */
562 __be32
563 nlmsvc_unlock(struct nlm_file *file, struct nlm_lock *lock)
564 {
565         int     error;
566
567         dprintk("lockd: nlmsvc_unlock(%s/%ld, pi=%d, %Ld-%Ld)\n",
568                                 file->f_file->f_path.dentry->d_inode->i_sb->s_id,
569                                 file->f_file->f_path.dentry->d_inode->i_ino,
570                                 lock->fl.fl_pid,
571                                 (long long)lock->fl.fl_start,
572                                 (long long)lock->fl.fl_end);
573
574         /* First, cancel any lock that might be there */
575         nlmsvc_cancel_blocked(file, lock);
576
577         lock->fl.fl_type = F_UNLCK;
578         error = vfs_lock_file(file->f_file, F_SETLK, &lock->fl, NULL);
579
580         return (error < 0)? nlm_lck_denied_nolocks : nlm_granted;
581 }
582
583 /*
584  * Cancel a previously blocked request.
585  *
586  * A cancel request always overrides any grant that may currently
587  * be in progress.
588  * The calling procedure must check whether the file can be closed.
589  */
590 __be32
591 nlmsvc_cancel_blocked(struct nlm_file *file, struct nlm_lock *lock)
592 {
593         struct nlm_block        *block;
594         int status = 0;
595
596         dprintk("lockd: nlmsvc_cancel(%s/%ld, pi=%d, %Ld-%Ld)\n",
597                                 file->f_file->f_path.dentry->d_inode->i_sb->s_id,
598                                 file->f_file->f_path.dentry->d_inode->i_ino,
599                                 lock->fl.fl_pid,
600                                 (long long)lock->fl.fl_start,
601                                 (long long)lock->fl.fl_end);
602
603         mutex_lock(&file->f_mutex);
604         block = nlmsvc_lookup_block(file, lock);
605         mutex_unlock(&file->f_mutex);
606         if (block != NULL) {
607                 vfs_cancel_lock(block->b_file->f_file,
608                                 &block->b_call->a_args.lock.fl);
609                 status = nlmsvc_unlink_block(block);
610                 nlmsvc_release_block(block);
611         }
612         return status ? nlm_lck_denied : nlm_granted;
613 }
614
615 /*
616  * This is a callback from the filesystem for VFS file lock requests.
617  * It will be used if fl_grant is defined and the filesystem can not
618  * respond to the request immediately.
619  * For GETLK request it will copy the reply to the nlm_block.
620  * For SETLK or SETLKW request it will get the local posix lock.
621  * In all cases it will move the block to the head of nlm_blocked q where
622  * nlmsvc_retry_blocked() can send back a reply for SETLKW or revisit the
623  * deferred rpc for GETLK and SETLK.
624  */
625 static void
626 nlmsvc_update_deferred_block(struct nlm_block *block, struct file_lock *conf,
627                              int result)
628 {
629         block->b_flags |= B_GOT_CALLBACK;
630         if (result == 0)
631                 block->b_granted = 1;
632         else
633                 block->b_flags |= B_TIMED_OUT;
634         if (conf) {
635                 if (block->b_fl)
636                         locks_copy_lock(block->b_fl, conf);
637         }
638 }
639
640 static int nlmsvc_grant_deferred(struct file_lock *fl, struct file_lock *conf,
641                                         int result)
642 {
643         struct nlm_block *block;
644         int rc = -ENOENT;
645
646         lock_kernel();
647         list_for_each_entry(block, &nlm_blocked, b_list) {
648                 if (nlm_compare_locks(&block->b_call->a_args.lock.fl, fl)) {
649                         dprintk("lockd: nlmsvc_notify_blocked block %p flags %d\n",
650                                                         block, block->b_flags);
651                         if (block->b_flags & B_QUEUED) {
652                                 if (block->b_flags & B_TIMED_OUT) {
653                                         rc = -ENOLCK;
654                                         break;
655                                 }
656                                 nlmsvc_update_deferred_block(block, conf, result);
657                         } else if (result == 0)
658                                 block->b_granted = 1;
659
660                         nlmsvc_insert_block(block, 0);
661                         svc_wake_up(block->b_daemon);
662                         rc = 0;
663                         break;
664                 }
665         }
666         unlock_kernel();
667         if (rc == -ENOENT)
668                 printk(KERN_WARNING "lockd: grant for unknown block\n");
669         return rc;
670 }
671
672 /*
673  * Unblock a blocked lock request. This is a callback invoked from the
674  * VFS layer when a lock on which we blocked is removed.
675  *
676  * This function doesn't grant the blocked lock instantly, but rather moves
677  * the block to the head of nlm_blocked where it can be picked up by lockd.
678  */
679 static void
680 nlmsvc_notify_blocked(struct file_lock *fl)
681 {
682         struct nlm_block        *block;
683
684         dprintk("lockd: VFS unblock notification for block %p\n", fl);
685         list_for_each_entry(block, &nlm_blocked, b_list) {
686                 if (nlm_compare_locks(&block->b_call->a_args.lock.fl, fl)) {
687                         nlmsvc_insert_block(block, 0);
688                         svc_wake_up(block->b_daemon);
689                         return;
690                 }
691         }
692
693         printk(KERN_WARNING "lockd: notification for unknown block!\n");
694 }
695
696 static int nlmsvc_same_owner(struct file_lock *fl1, struct file_lock *fl2)
697 {
698         return fl1->fl_owner == fl2->fl_owner && fl1->fl_pid == fl2->fl_pid;
699 }
700
701 struct lock_manager_operations nlmsvc_lock_operations = {
702         .fl_compare_owner = nlmsvc_same_owner,
703         .fl_notify = nlmsvc_notify_blocked,
704         .fl_grant = nlmsvc_grant_deferred,
705 };
706
707 /*
708  * Try to claim a lock that was previously blocked.
709  *
710  * Note that we use both the RPC_GRANTED_MSG call _and_ an async
711  * RPC thread when notifying the client. This seems like overkill...
712  * Here's why:
713  *  -   we don't want to use a synchronous RPC thread, otherwise
714  *      we might find ourselves hanging on a dead portmapper.
715  *  -   Some lockd implementations (e.g. HP) don't react to
716  *      RPC_GRANTED calls; they seem to insist on RPC_GRANTED_MSG calls.
717  */
718 static void
719 nlmsvc_grant_blocked(struct nlm_block *block)
720 {
721         struct nlm_file         *file = block->b_file;
722         struct nlm_lock         *lock = &block->b_call->a_args.lock;
723         int                     error;
724
725         dprintk("lockd: grant blocked lock %p\n", block);
726
727         kref_get(&block->b_count);
728
729         /* Unlink block request from list */
730         nlmsvc_unlink_block(block);
731
732         /* If b_granted is true this means we've been here before.
733          * Just retry the grant callback, possibly refreshing the RPC
734          * binding */
735         if (block->b_granted) {
736                 nlm_rebind_host(block->b_host);
737                 goto callback;
738         }
739
740         /* Try the lock operation again */
741         lock->fl.fl_flags |= FL_SLEEP;
742         error = vfs_lock_file(file->f_file, F_SETLK, &lock->fl, NULL);
743         lock->fl.fl_flags &= ~FL_SLEEP;
744
745         switch (error) {
746         case 0:
747                 break;
748         case -EAGAIN:
749         case -EINPROGRESS:
750                 dprintk("lockd: lock still blocked error %d\n", error);
751                 nlmsvc_insert_block(block, NLM_NEVER);
752                 nlmsvc_release_block(block);
753                 return;
754         default:
755                 printk(KERN_WARNING "lockd: unexpected error %d in %s!\n",
756                                 -error, __FUNCTION__);
757                 nlmsvc_insert_block(block, 10 * HZ);
758                 nlmsvc_release_block(block);
759                 return;
760         }
761
762 callback:
763         /* Lock was granted by VFS. */
764         dprintk("lockd: GRANTing blocked lock.\n");
765         block->b_granted = 1;
766
767         /* keep block on the list, but don't reattempt until the RPC
768          * completes or the submission fails
769          */
770         nlmsvc_insert_block(block, NLM_NEVER);
771
772         /* Call the client -- use a soft RPC task since nlmsvc_retry_blocked
773          * will queue up a new one if this one times out
774          */
775         error = nlm_async_call(block->b_call, NLMPROC_GRANTED_MSG,
776                                 &nlmsvc_grant_ops);
777
778         /* RPC submission failed, wait a bit and retry */
779         if (error < 0)
780                 nlmsvc_insert_block(block, 10 * HZ);
781 }
782
783 /*
784  * This is the callback from the RPC layer when the NLM_GRANTED_MSG
785  * RPC call has succeeded or timed out.
786  * Like all RPC callbacks, it is invoked by the rpciod process, so it
787  * better not sleep. Therefore, we put the blocked lock on the nlm_blocked
788  * chain once more in order to have it removed by lockd itself (which can
789  * then sleep on the file semaphore without disrupting e.g. the nfs client).
790  */
791 static void nlmsvc_grant_callback(struct rpc_task *task, void *data)
792 {
793         struct nlm_rqst         *call = data;
794         struct nlm_block        *block = call->a_block;
795         unsigned long           timeout;
796
797         dprintk("lockd: GRANT_MSG RPC callback\n");
798
799         /* if the block is not on a list at this point then it has
800          * been invalidated. Don't try to requeue it.
801          *
802          * FIXME: it's possible that the block is removed from the list
803          * after this check but before the nlmsvc_insert_block. In that
804          * case it will be added back. Perhaps we need better locking
805          * for nlm_blocked?
806          */
807         if (list_empty(&block->b_list))
808                 return;
809
810         /* Technically, we should down the file semaphore here. Since we
811          * move the block towards the head of the queue only, no harm
812          * can be done, though. */
813         if (task->tk_status < 0) {
814                 /* RPC error: Re-insert for retransmission */
815                 timeout = 10 * HZ;
816         } else {
817                 /* Call was successful, now wait for client callback */
818                 timeout = 60 * HZ;
819         }
820         nlmsvc_insert_block(block, timeout);
821         svc_wake_up(block->b_daemon);
822 }
823
824 static void nlmsvc_grant_release(void *data)
825 {
826         struct nlm_rqst         *call = data;
827
828         nlmsvc_release_block(call->a_block);
829 }
830
831 static const struct rpc_call_ops nlmsvc_grant_ops = {
832         .rpc_call_done = nlmsvc_grant_callback,
833         .rpc_release = nlmsvc_grant_release,
834 };
835
836 /*
837  * We received a GRANT_RES callback. Try to find the corresponding
838  * block.
839  */
840 void
841 nlmsvc_grant_reply(struct nlm_cookie *cookie, __be32 status)
842 {
843         struct nlm_block        *block;
844
845         dprintk("grant_reply: looking for cookie %x, s=%d \n",
846                 *(unsigned int *)(cookie->data), status);
847         if (!(block = nlmsvc_find_block(cookie)))
848                 return;
849
850         if (block) {
851                 if (status == nlm_lck_denied_grace_period) {
852                         /* Try again in a couple of seconds */
853                         nlmsvc_insert_block(block, 10 * HZ);
854                 } else {
855                         /* Lock is now held by client, or has been rejected.
856                          * In both cases, the block should be removed. */
857                         nlmsvc_unlink_block(block);
858                 }
859         }
860         nlmsvc_release_block(block);
861 }
862
863 /* Helper function to handle retry of a deferred block.
864  * If it is a blocking lock, call grant_blocked.
865  * For a non-blocking lock or test lock, revisit the request.
866  */
867 static void
868 retry_deferred_block(struct nlm_block *block)
869 {
870         if (!(block->b_flags & B_GOT_CALLBACK))
871                 block->b_flags |= B_TIMED_OUT;
872         nlmsvc_insert_block(block, NLM_TIMEOUT);
873         dprintk("revisit block %p flags %d\n",  block, block->b_flags);
874         if (block->b_deferred_req) {
875                 block->b_deferred_req->revisit(block->b_deferred_req, 0);
876                 block->b_deferred_req = NULL;
877         }
878 }
879
880 /*
881  * Retry all blocked locks that have been notified. This is where lockd
882  * picks up locks that can be granted, or grant notifications that must
883  * be retransmitted.
884  */
885 unsigned long
886 nlmsvc_retry_blocked(void)
887 {
888         unsigned long   timeout = MAX_SCHEDULE_TIMEOUT;
889         struct nlm_block *block;
890
891         while (!list_empty(&nlm_blocked) && !kthread_should_stop()) {
892                 block = list_entry(nlm_blocked.next, struct nlm_block, b_list);
893
894                 if (block->b_when == NLM_NEVER)
895                         break;
896                 if (time_after(block->b_when,jiffies)) {
897                         timeout = block->b_when - jiffies;
898                         break;
899                 }
900
901                 dprintk("nlmsvc_retry_blocked(%p, when=%ld)\n",
902                         block, block->b_when);
903                 if (block->b_flags & B_QUEUED) {
904                         dprintk("nlmsvc_retry_blocked delete block (%p, granted=%d, flags=%d)\n",
905                                 block, block->b_granted, block->b_flags);
906                         retry_deferred_block(block);
907                 } else
908                         nlmsvc_grant_blocked(block);
909         }
910
911         return timeout;
912 }