Merge branch 'for-linus' of git://git390.marist.edu/pub/scm/linux-2.6
[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         nlm_get_host(host);
184         call = nlm_alloc_call(host);
185         if (call == NULL)
186                 return NULL;
187
188         /* Allocate memory for block, and initialize arguments */
189         block = kzalloc(sizeof(*block), GFP_KERNEL);
190         if (block == NULL)
191                 goto failed;
192         kref_init(&block->b_count);
193         INIT_LIST_HEAD(&block->b_list);
194         INIT_LIST_HEAD(&block->b_flist);
195
196         if (!nlmsvc_setgrantargs(call, lock))
197                 goto failed_free;
198
199         /* Set notifier function for VFS, and init args */
200         call->a_args.lock.fl.fl_flags |= FL_SLEEP;
201         call->a_args.lock.fl.fl_lmops = &nlmsvc_lock_operations;
202         nlmclnt_next_cookie(&call->a_args.cookie);
203
204         dprintk("lockd: created block %p...\n", block);
205
206         /* Create and initialize the block */
207         block->b_daemon = rqstp->rq_server;
208         block->b_host   = host;
209         block->b_file   = file;
210         block->b_fl = NULL;
211         file->f_count++;
212
213         /* Add to file's list of blocks */
214         list_add(&block->b_flist, &file->f_blocks);
215
216         /* Set up RPC arguments for callback */
217         block->b_call = call;
218         call->a_flags   = RPC_TASK_ASYNC;
219         call->a_block = block;
220
221         return block;
222
223 failed_free:
224         kfree(block);
225 failed:
226         nlm_release_call(call);
227         return NULL;
228 }
229
230 /*
231  * Delete a block.
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_host *host, struct nlm_lock *lock, int wait,
363             struct nlm_cookie *cookie, int reclaim)
364 {
365         struct nlm_block        *block = NULL;
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         /* Lock file against concurrent access */
378         mutex_lock(&file->f_mutex);
379         /* Get existing block (in case client is busy-waiting)
380          * or create new block
381          */
382         block = nlmsvc_lookup_block(file, lock);
383         if (block == NULL) {
384                 block = nlmsvc_create_block(rqstp, host, file, lock, cookie);
385                 ret = nlm_lck_denied_nolocks;
386                 if (block == NULL)
387                         goto out;
388                 lock = &block->b_call->a_args.lock;
389         } else
390                 lock->fl.fl_flags &= ~FL_SLEEP;
391
392         if (block->b_flags & B_QUEUED) {
393                 dprintk("lockd: nlmsvc_lock deferred block %p flags %d\n",
394                                                         block, block->b_flags);
395                 if (block->b_granted) {
396                         nlmsvc_unlink_block(block);
397                         ret = nlm_granted;
398                         goto out;
399                 }
400                 if (block->b_flags & B_TIMED_OUT) {
401                         nlmsvc_unlink_block(block);
402                         ret = nlm_lck_denied;
403                         goto out;
404                 }
405                 ret = nlm_drop_reply;
406                 goto out;
407         }
408
409         if (locks_in_grace() && !reclaim) {
410                 ret = nlm_lck_denied_grace_period;
411                 goto out;
412         }
413         if (reclaim && !locks_in_grace()) {
414                 ret = nlm_lck_denied_grace_period;
415                 goto out;
416         }
417
418         if (!wait)
419                 lock->fl.fl_flags &= ~FL_SLEEP;
420         error = vfs_lock_file(file->f_file, F_SETLK, &lock->fl, NULL);
421         lock->fl.fl_flags &= ~FL_SLEEP;
422
423         dprintk("lockd: vfs_lock_file returned %d\n", error);
424         switch (error) {
425                 case 0:
426                         ret = nlm_granted;
427                         goto out;
428                 case -EAGAIN:
429                         /*
430                          * If this is a blocking request for an
431                          * already pending lock request then we need
432                          * to put it back on lockd's block list
433                          */
434                         if (wait)
435                                 break;
436                         ret = nlm_lck_denied;
437                         goto out;
438                 case FILE_LOCK_DEFERRED:
439                         if (wait)
440                                 break;
441                         /* Filesystem lock operation is in progress
442                            Add it to the queue waiting for callback */
443                         ret = nlmsvc_defer_lock_rqst(rqstp, block);
444                         goto out;
445                 case -EDEADLK:
446                         ret = nlm_deadlock;
447                         goto out;
448                 default:                        /* includes ENOLCK */
449                         ret = nlm_lck_denied_nolocks;
450                         goto out;
451         }
452
453         ret = nlm_lck_blocked;
454
455         /* Append to list of blocked */
456         nlmsvc_insert_block(block, NLM_NEVER);
457 out:
458         mutex_unlock(&file->f_mutex);
459         nlmsvc_release_block(block);
460         dprintk("lockd: nlmsvc_lock returned %u\n", ret);
461         return ret;
462 }
463
464 /*
465  * Test for presence of a conflicting lock.
466  */
467 __be32
468 nlmsvc_testlock(struct svc_rqst *rqstp, struct nlm_file *file,
469                 struct nlm_host *host, struct nlm_lock *lock,
470                 struct nlm_lock *conflock, struct nlm_cookie *cookie)
471 {
472         struct nlm_block        *block = NULL;
473         int                     error;
474         __be32                  ret;
475
476         dprintk("lockd: nlmsvc_testlock(%s/%ld, ty=%d, %Ld-%Ld)\n",
477                                 file->f_file->f_path.dentry->d_inode->i_sb->s_id,
478                                 file->f_file->f_path.dentry->d_inode->i_ino,
479                                 lock->fl.fl_type,
480                                 (long long)lock->fl.fl_start,
481                                 (long long)lock->fl.fl_end);
482
483         /* Get existing block (in case client is busy-waiting) */
484         block = nlmsvc_lookup_block(file, lock);
485
486         if (block == NULL) {
487                 struct file_lock *conf = kzalloc(sizeof(*conf), GFP_KERNEL);
488
489                 if (conf == NULL)
490                         return nlm_granted;
491                 block = nlmsvc_create_block(rqstp, host, file, lock, cookie);
492                 if (block == NULL) {
493                         kfree(conf);
494                         return nlm_granted;
495                 }
496                 block->b_fl = conf;
497         }
498         if (block->b_flags & B_QUEUED) {
499                 dprintk("lockd: nlmsvc_testlock deferred block %p flags %d fl %p\n",
500                         block, block->b_flags, block->b_fl);
501                 if (block->b_flags & B_TIMED_OUT) {
502                         nlmsvc_unlink_block(block);
503                         ret = nlm_lck_denied;
504                         goto out;
505                 }
506                 if (block->b_flags & B_GOT_CALLBACK) {
507                         nlmsvc_unlink_block(block);
508                         if (block->b_fl != NULL
509                                         && block->b_fl->fl_type != F_UNLCK) {
510                                 lock->fl = *block->b_fl;
511                                 goto conf_lock;
512                         } else {
513                                 ret = nlm_granted;
514                                 goto out;
515                         }
516                 }
517                 ret = nlm_drop_reply;
518                 goto out;
519         }
520
521         if (locks_in_grace()) {
522                 ret = nlm_lck_denied_grace_period;
523                 goto out;
524         }
525         error = vfs_test_lock(file->f_file, &lock->fl);
526         if (error == FILE_LOCK_DEFERRED) {
527                 ret = nlmsvc_defer_lock_rqst(rqstp, block);
528                 goto out;
529         }
530         if (error) {
531                 ret = nlm_lck_denied_nolocks;
532                 goto out;
533         }
534         if (lock->fl.fl_type == F_UNLCK) {
535                 ret = nlm_granted;
536                 goto out;
537         }
538
539 conf_lock:
540         dprintk("lockd: conflicting lock(ty=%d, %Ld-%Ld)\n",
541                 lock->fl.fl_type, (long long)lock->fl.fl_start,
542                 (long long)lock->fl.fl_end);
543         conflock->caller = "somehost";  /* FIXME */
544         conflock->len = strlen(conflock->caller);
545         conflock->oh.len = 0;           /* don't return OH info */
546         conflock->svid = lock->fl.fl_pid;
547         conflock->fl.fl_type = lock->fl.fl_type;
548         conflock->fl.fl_start = lock->fl.fl_start;
549         conflock->fl.fl_end = lock->fl.fl_end;
550         ret = nlm_lck_denied;
551 out:
552         if (block)
553                 nlmsvc_release_block(block);
554         return ret;
555 }
556
557 /*
558  * Remove a lock.
559  * This implies a CANCEL call: We send a GRANT_MSG, the client replies
560  * with a GRANT_RES call which gets lost, and calls UNLOCK immediately
561  * afterwards. In this case the block will still be there, and hence
562  * must be removed.
563  */
564 __be32
565 nlmsvc_unlock(struct nlm_file *file, struct nlm_lock *lock)
566 {
567         int     error;
568
569         dprintk("lockd: nlmsvc_unlock(%s/%ld, pi=%d, %Ld-%Ld)\n",
570                                 file->f_file->f_path.dentry->d_inode->i_sb->s_id,
571                                 file->f_file->f_path.dentry->d_inode->i_ino,
572                                 lock->fl.fl_pid,
573                                 (long long)lock->fl.fl_start,
574                                 (long long)lock->fl.fl_end);
575
576         /* First, cancel any lock that might be there */
577         nlmsvc_cancel_blocked(file, lock);
578
579         lock->fl.fl_type = F_UNLCK;
580         error = vfs_lock_file(file->f_file, F_SETLK, &lock->fl, NULL);
581
582         return (error < 0)? nlm_lck_denied_nolocks : nlm_granted;
583 }
584
585 /*
586  * Cancel a previously blocked request.
587  *
588  * A cancel request always overrides any grant that may currently
589  * be in progress.
590  * The calling procedure must check whether the file can be closed.
591  */
592 __be32
593 nlmsvc_cancel_blocked(struct nlm_file *file, struct nlm_lock *lock)
594 {
595         struct nlm_block        *block;
596         int status = 0;
597
598         dprintk("lockd: nlmsvc_cancel(%s/%ld, pi=%d, %Ld-%Ld)\n",
599                                 file->f_file->f_path.dentry->d_inode->i_sb->s_id,
600                                 file->f_file->f_path.dentry->d_inode->i_ino,
601                                 lock->fl.fl_pid,
602                                 (long long)lock->fl.fl_start,
603                                 (long long)lock->fl.fl_end);
604
605         if (locks_in_grace())
606                 return nlm_lck_denied_grace_period;
607
608         mutex_lock(&file->f_mutex);
609         block = nlmsvc_lookup_block(file, lock);
610         mutex_unlock(&file->f_mutex);
611         if (block != NULL) {
612                 vfs_cancel_lock(block->b_file->f_file,
613                                 &block->b_call->a_args.lock.fl);
614                 status = nlmsvc_unlink_block(block);
615                 nlmsvc_release_block(block);
616         }
617         return status ? nlm_lck_denied : nlm_granted;
618 }
619
620 /*
621  * This is a callback from the filesystem for VFS file lock requests.
622  * It will be used if fl_grant is defined and the filesystem can not
623  * respond to the request immediately.
624  * For GETLK request it will copy the reply to the nlm_block.
625  * For SETLK or SETLKW request it will get the local posix lock.
626  * In all cases it will move the block to the head of nlm_blocked q where
627  * nlmsvc_retry_blocked() can send back a reply for SETLKW or revisit the
628  * deferred rpc for GETLK and SETLK.
629  */
630 static void
631 nlmsvc_update_deferred_block(struct nlm_block *block, struct file_lock *conf,
632                              int result)
633 {
634         block->b_flags |= B_GOT_CALLBACK;
635         if (result == 0)
636                 block->b_granted = 1;
637         else
638                 block->b_flags |= B_TIMED_OUT;
639         if (conf) {
640                 if (block->b_fl)
641                         __locks_copy_lock(block->b_fl, conf);
642         }
643 }
644
645 static int nlmsvc_grant_deferred(struct file_lock *fl, struct file_lock *conf,
646                                         int result)
647 {
648         struct nlm_block *block;
649         int rc = -ENOENT;
650
651         lock_kernel();
652         list_for_each_entry(block, &nlm_blocked, b_list) {
653                 if (nlm_compare_locks(&block->b_call->a_args.lock.fl, fl)) {
654                         dprintk("lockd: nlmsvc_notify_blocked block %p flags %d\n",
655                                                         block, block->b_flags);
656                         if (block->b_flags & B_QUEUED) {
657                                 if (block->b_flags & B_TIMED_OUT) {
658                                         rc = -ENOLCK;
659                                         break;
660                                 }
661                                 nlmsvc_update_deferred_block(block, conf, result);
662                         } else if (result == 0)
663                                 block->b_granted = 1;
664
665                         nlmsvc_insert_block(block, 0);
666                         svc_wake_up(block->b_daemon);
667                         rc = 0;
668                         break;
669                 }
670         }
671         unlock_kernel();
672         if (rc == -ENOENT)
673                 printk(KERN_WARNING "lockd: grant for unknown block\n");
674         return rc;
675 }
676
677 /*
678  * Unblock a blocked lock request. This is a callback invoked from the
679  * VFS layer when a lock on which we blocked is removed.
680  *
681  * This function doesn't grant the blocked lock instantly, but rather moves
682  * the block to the head of nlm_blocked where it can be picked up by lockd.
683  */
684 static void
685 nlmsvc_notify_blocked(struct file_lock *fl)
686 {
687         struct nlm_block        *block;
688
689         dprintk("lockd: VFS unblock notification for block %p\n", fl);
690         list_for_each_entry(block, &nlm_blocked, b_list) {
691                 if (nlm_compare_locks(&block->b_call->a_args.lock.fl, fl)) {
692                         nlmsvc_insert_block(block, 0);
693                         svc_wake_up(block->b_daemon);
694                         return;
695                 }
696         }
697
698         printk(KERN_WARNING "lockd: notification for unknown block!\n");
699 }
700
701 static int nlmsvc_same_owner(struct file_lock *fl1, struct file_lock *fl2)
702 {
703         return fl1->fl_owner == fl2->fl_owner && fl1->fl_pid == fl2->fl_pid;
704 }
705
706 struct lock_manager_operations nlmsvc_lock_operations = {
707         .fl_compare_owner = nlmsvc_same_owner,
708         .fl_notify = nlmsvc_notify_blocked,
709         .fl_grant = nlmsvc_grant_deferred,
710 };
711
712 /*
713  * Try to claim a lock that was previously blocked.
714  *
715  * Note that we use both the RPC_GRANTED_MSG call _and_ an async
716  * RPC thread when notifying the client. This seems like overkill...
717  * Here's why:
718  *  -   we don't want to use a synchronous RPC thread, otherwise
719  *      we might find ourselves hanging on a dead portmapper.
720  *  -   Some lockd implementations (e.g. HP) don't react to
721  *      RPC_GRANTED calls; they seem to insist on RPC_GRANTED_MSG calls.
722  */
723 static void
724 nlmsvc_grant_blocked(struct nlm_block *block)
725 {
726         struct nlm_file         *file = block->b_file;
727         struct nlm_lock         *lock = &block->b_call->a_args.lock;
728         int                     error;
729
730         dprintk("lockd: grant blocked lock %p\n", block);
731
732         kref_get(&block->b_count);
733
734         /* Unlink block request from list */
735         nlmsvc_unlink_block(block);
736
737         /* If b_granted is true this means we've been here before.
738          * Just retry the grant callback, possibly refreshing the RPC
739          * binding */
740         if (block->b_granted) {
741                 nlm_rebind_host(block->b_host);
742                 goto callback;
743         }
744
745         /* Try the lock operation again */
746         lock->fl.fl_flags |= FL_SLEEP;
747         error = vfs_lock_file(file->f_file, F_SETLK, &lock->fl, NULL);
748         lock->fl.fl_flags &= ~FL_SLEEP;
749
750         switch (error) {
751         case 0:
752                 break;
753         case FILE_LOCK_DEFERRED:
754                 dprintk("lockd: lock still blocked error %d\n", error);
755                 nlmsvc_insert_block(block, NLM_NEVER);
756                 nlmsvc_release_block(block);
757                 return;
758         default:
759                 printk(KERN_WARNING "lockd: unexpected error %d in %s!\n",
760                                 -error, __func__);
761                 nlmsvc_insert_block(block, 10 * HZ);
762                 nlmsvc_release_block(block);
763                 return;
764         }
765
766 callback:
767         /* Lock was granted by VFS. */
768         dprintk("lockd: GRANTing blocked lock.\n");
769         block->b_granted = 1;
770
771         /* keep block on the list, but don't reattempt until the RPC
772          * completes or the submission fails
773          */
774         nlmsvc_insert_block(block, NLM_NEVER);
775
776         /* Call the client -- use a soft RPC task since nlmsvc_retry_blocked
777          * will queue up a new one if this one times out
778          */
779         error = nlm_async_call(block->b_call, NLMPROC_GRANTED_MSG,
780                                 &nlmsvc_grant_ops);
781
782         /* RPC submission failed, wait a bit and retry */
783         if (error < 0)
784                 nlmsvc_insert_block(block, 10 * HZ);
785 }
786
787 /*
788  * This is the callback from the RPC layer when the NLM_GRANTED_MSG
789  * RPC call has succeeded or timed out.
790  * Like all RPC callbacks, it is invoked by the rpciod process, so it
791  * better not sleep. Therefore, we put the blocked lock on the nlm_blocked
792  * chain once more in order to have it removed by lockd itself (which can
793  * then sleep on the file semaphore without disrupting e.g. the nfs client).
794  */
795 static void nlmsvc_grant_callback(struct rpc_task *task, void *data)
796 {
797         struct nlm_rqst         *call = data;
798         struct nlm_block        *block = call->a_block;
799         unsigned long           timeout;
800
801         dprintk("lockd: GRANT_MSG RPC callback\n");
802
803         lock_kernel();
804         /* if the block is not on a list at this point then it has
805          * been invalidated. Don't try to requeue it.
806          *
807          * FIXME: it's possible that the block is removed from the list
808          * after this check but before the nlmsvc_insert_block. In that
809          * case it will be added back. Perhaps we need better locking
810          * for nlm_blocked?
811          */
812         if (list_empty(&block->b_list))
813                 goto out;
814
815         /* Technically, we should down the file semaphore here. Since we
816          * move the block towards the head of the queue only, no harm
817          * can be done, though. */
818         if (task->tk_status < 0) {
819                 /* RPC error: Re-insert for retransmission */
820                 timeout = 10 * HZ;
821         } else {
822                 /* Call was successful, now wait for client callback */
823                 timeout = 60 * HZ;
824         }
825         nlmsvc_insert_block(block, timeout);
826         svc_wake_up(block->b_daemon);
827 out:
828         unlock_kernel();
829 }
830
831 static void nlmsvc_grant_release(void *data)
832 {
833         struct nlm_rqst         *call = data;
834
835         lock_kernel();
836         nlmsvc_release_block(call->a_block);
837         unlock_kernel();
838 }
839
840 static const struct rpc_call_ops nlmsvc_grant_ops = {
841         .rpc_call_done = nlmsvc_grant_callback,
842         .rpc_release = nlmsvc_grant_release,
843 };
844
845 /*
846  * We received a GRANT_RES callback. Try to find the corresponding
847  * block.
848  */
849 void
850 nlmsvc_grant_reply(struct nlm_cookie *cookie, __be32 status)
851 {
852         struct nlm_block        *block;
853
854         dprintk("grant_reply: looking for cookie %x, s=%d \n",
855                 *(unsigned int *)(cookie->data), status);
856         if (!(block = nlmsvc_find_block(cookie)))
857                 return;
858
859         if (block) {
860                 if (status == nlm_lck_denied_grace_period) {
861                         /* Try again in a couple of seconds */
862                         nlmsvc_insert_block(block, 10 * HZ);
863                 } else {
864                         /* Lock is now held by client, or has been rejected.
865                          * In both cases, the block should be removed. */
866                         nlmsvc_unlink_block(block);
867                 }
868         }
869         nlmsvc_release_block(block);
870 }
871
872 /* Helper function to handle retry of a deferred block.
873  * If it is a blocking lock, call grant_blocked.
874  * For a non-blocking lock or test lock, revisit the request.
875  */
876 static void
877 retry_deferred_block(struct nlm_block *block)
878 {
879         if (!(block->b_flags & B_GOT_CALLBACK))
880                 block->b_flags |= B_TIMED_OUT;
881         nlmsvc_insert_block(block, NLM_TIMEOUT);
882         dprintk("revisit block %p flags %d\n",  block, block->b_flags);
883         if (block->b_deferred_req) {
884                 block->b_deferred_req->revisit(block->b_deferred_req, 0);
885                 block->b_deferred_req = NULL;
886         }
887 }
888
889 /*
890  * Retry all blocked locks that have been notified. This is where lockd
891  * picks up locks that can be granted, or grant notifications that must
892  * be retransmitted.
893  */
894 unsigned long
895 nlmsvc_retry_blocked(void)
896 {
897         unsigned long   timeout = MAX_SCHEDULE_TIMEOUT;
898         struct nlm_block *block;
899
900         while (!list_empty(&nlm_blocked) && !kthread_should_stop()) {
901                 block = list_entry(nlm_blocked.next, struct nlm_block, b_list);
902
903                 if (block->b_when == NLM_NEVER)
904                         break;
905                 if (time_after(block->b_when, jiffies)) {
906                         timeout = block->b_when - jiffies;
907                         break;
908                 }
909
910                 dprintk("nlmsvc_retry_blocked(%p, when=%ld)\n",
911                         block, block->b_when);
912                 if (block->b_flags & B_QUEUED) {
913                         dprintk("nlmsvc_retry_blocked delete block (%p, granted=%d, flags=%d)\n",
914                                 block, block->b_granted, block->b_flags);
915                         retry_deferred_block(block);
916                 } else
917                         nlmsvc_grant_blocked(block);
918         }
919
920         return timeout;
921 }