initial commit with v2.6.9
[linux-2.6.9-moxart.git] / fs / lockd / svclock.c
blob80e7cc4d44a2e15a546df7de18af8a0e3de9c680
1 /*
2 * linux/fs/lockd/svclock.c
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.
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.
20 * Copyright (C) 1996, Olaf Kirch <okir@monad.swb.de>
23 #include <linux/config.h>
24 #include <linux/types.h>
25 #include <linux/errno.h>
26 #include <linux/kernel.h>
27 #include <linux/sched.h>
28 #include <linux/smp_lock.h>
29 #include <linux/sunrpc/clnt.h>
30 #include <linux/sunrpc/svc.h>
31 #include <linux/lockd/nlm.h>
32 #include <linux/lockd/lockd.h>
34 #define NLMDBG_FACILITY NLMDBG_SVCLOCK
36 #ifdef CONFIG_LOCKD_V4
37 #define nlm_deadlock nlm4_deadlock
38 #else
39 #define nlm_deadlock nlm_lck_denied
40 #endif
42 static void nlmsvc_insert_block(struct nlm_block *block, unsigned long);
43 static int nlmsvc_remove_block(struct nlm_block *block);
44 static void nlmsvc_grant_callback(struct rpc_task *task);
47 * The list of blocked locks to retry
49 static struct nlm_block * nlm_blocked;
52 * Insert a blocked lock into the global list
54 static void
55 nlmsvc_insert_block(struct nlm_block *block, unsigned long when)
57 struct nlm_block **bp, *b;
59 dprintk("lockd: nlmsvc_insert_block(%p, %ld)\n", block, when);
60 if (block->b_queued)
61 nlmsvc_remove_block(block);
62 bp = &nlm_blocked;
63 if (when != NLM_NEVER) {
64 if ((when += jiffies) == NLM_NEVER)
65 when ++;
66 while ((b = *bp) && time_before_eq(b->b_when,when) && b->b_when != NLM_NEVER)
67 bp = &b->b_next;
68 } else
69 while ((b = *bp) != 0)
70 bp = &b->b_next;
72 block->b_queued = 1;
73 block->b_when = when;
74 block->b_next = b;
75 *bp = block;
79 * Remove a block from the global list
81 static int
82 nlmsvc_remove_block(struct nlm_block *block)
84 struct nlm_block **bp, *b;
86 if (!block->b_queued)
87 return 1;
88 for (bp = &nlm_blocked; (b = *bp) != 0; bp = &b->b_next) {
89 if (b == block) {
90 *bp = block->b_next;
91 block->b_queued = 0;
92 return 1;
96 return 0;
100 * Find a block for a given lock and optionally remove it from
101 * the list.
103 static struct nlm_block *
104 nlmsvc_lookup_block(struct nlm_file *file, struct nlm_lock *lock, int remove)
106 struct nlm_block **head, *block;
107 struct file_lock *fl;
109 dprintk("lockd: nlmsvc_lookup_block f=%p pd=%d %Ld-%Ld ty=%d\n",
110 file, lock->fl.fl_pid,
111 (long long)lock->fl.fl_start,
112 (long long)lock->fl.fl_end, lock->fl.fl_type);
113 for (head = &nlm_blocked; (block = *head) != 0; head = &block->b_next) {
114 fl = &block->b_call.a_args.lock.fl;
115 dprintk("lockd: check f=%p pd=%d %Ld-%Ld ty=%d cookie=%x\n",
116 block->b_file, fl->fl_pid,
117 (long long)fl->fl_start,
118 (long long)fl->fl_end, fl->fl_type,
119 *(unsigned int*)(block->b_call.a_args.cookie.data));
120 if (block->b_file == file && nlm_compare_locks(fl, &lock->fl)) {
121 if (remove) {
122 *head = block->b_next;
123 block->b_queued = 0;
125 return block;
129 return NULL;
132 static inline int nlm_cookie_match(struct nlm_cookie *a, struct nlm_cookie *b)
134 if(a->len != b->len)
135 return 0;
136 if(memcmp(a->data,b->data,a->len))
137 return 0;
138 return 1;
142 * Find a block with a given NLM cookie.
144 static inline struct nlm_block *
145 nlmsvc_find_block(struct nlm_cookie *cookie, struct sockaddr_in *sin)
147 struct nlm_block *block;
149 for (block = nlm_blocked; block; block = block->b_next) {
150 dprintk("cookie: head of blocked queue %p, block %p\n",
151 nlm_blocked, block);
152 if (nlm_cookie_match(&block->b_call.a_args.cookie,cookie)
153 && nlm_cmp_addr(sin, &block->b_host->h_addr))
154 break;
157 return block;
161 * Create a block and initialize it.
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.
170 static inline struct nlm_block *
171 nlmsvc_create_block(struct svc_rqst *rqstp, struct nlm_file *file,
172 struct nlm_lock *lock, struct nlm_cookie *cookie)
174 struct nlm_block *block;
175 struct nlm_host *host;
176 struct nlm_rqst *call;
178 /* Create host handle for callback */
179 host = nlmclnt_lookup_host(&rqstp->rq_addr,
180 rqstp->rq_prot, rqstp->rq_vers);
181 if (host == NULL)
182 return NULL;
184 /* Allocate memory for block, and initialize arguments */
185 if (!(block = (struct nlm_block *) kmalloc(sizeof(*block), GFP_KERNEL)))
186 goto failed;
187 memset(block, 0, sizeof(*block));
188 locks_init_lock(&block->b_call.a_args.lock.fl);
189 locks_init_lock(&block->b_call.a_res.lock.fl);
191 if (!nlmclnt_setgrantargs(&block->b_call, lock))
192 goto failed_free;
194 /* Set notifier function for VFS, and init args */
195 block->b_call.a_args.lock.fl.fl_lmops = &nlmsvc_lock_operations;
196 block->b_call.a_args.cookie = *cookie; /* see above */
198 dprintk("lockd: created block %p...\n", block);
200 /* Create and initialize the block */
201 block->b_daemon = rqstp->rq_server;
202 block->b_host = host;
203 block->b_file = file;
205 /* Add to file's list of blocks */
206 block->b_fnext = file->f_blocks;
207 file->f_blocks = block;
209 /* Set up RPC arguments for callback */
210 call = &block->b_call;
211 call->a_host = host;
212 call->a_flags = RPC_TASK_ASYNC;
214 return block;
216 failed_free:
217 kfree(block);
218 failed:
219 nlm_release_host(host);
220 return NULL;
224 * Delete a block. If the lock was cancelled or the grant callback
225 * failed, unlock is set to 1.
226 * It is the caller's responsibility to check whether the file
227 * can be closed hereafter.
229 static void
230 nlmsvc_delete_block(struct nlm_block *block, int unlock)
232 struct file_lock *fl = &block->b_call.a_args.lock.fl;
233 struct nlm_file *file = block->b_file;
234 struct nlm_block **bp;
236 dprintk("lockd: deleting block %p...\n", block);
238 /* Remove block from list */
239 nlmsvc_remove_block(block);
240 if (fl->fl_next)
241 posix_unblock_lock(file->f_file, fl);
242 if (unlock) {
243 fl->fl_type = F_UNLCK;
244 posix_lock_file(file->f_file, fl);
245 block->b_granted = 0;
248 /* If the block is in the middle of a GRANT callback,
249 * don't kill it yet. */
250 if (block->b_incall) {
251 nlmsvc_insert_block(block, NLM_NEVER);
252 block->b_done = 1;
253 return;
256 /* Remove block from file's list of blocks */
257 for (bp = &file->f_blocks; *bp; bp = &(*bp)->b_fnext) {
258 if (*bp == block) {
259 *bp = block->b_fnext;
260 break;
264 if (block->b_host)
265 nlm_release_host(block->b_host);
266 nlmclnt_freegrantargs(&block->b_call);
267 kfree(block);
271 * Loop over all blocks and perform the action specified.
272 * (NLM_ACT_CHECK handled by nlmsvc_inspect_file).
275 nlmsvc_traverse_blocks(struct nlm_host *host, struct nlm_file *file, int action)
277 struct nlm_block *block, *next;
279 down(&file->f_sema);
280 for (block = file->f_blocks; block; block = next) {
281 next = block->b_fnext;
282 if (action == NLM_ACT_MARK)
283 block->b_host->h_inuse = 1;
284 else if (action == NLM_ACT_UNLOCK) {
285 if (host == NULL || host == block->b_host)
286 nlmsvc_delete_block(block, 1);
289 up(&file->f_sema);
290 return 0;
294 * Attempt to establish a lock, and if it can't be granted, block it
295 * if required.
298 nlmsvc_lock(struct svc_rqst *rqstp, struct nlm_file *file,
299 struct nlm_lock *lock, int wait, struct nlm_cookie *cookie)
301 struct file_lock *conflock;
302 struct nlm_block *block;
303 int error;
305 dprintk("lockd: nlmsvc_lock(%s/%ld, ty=%d, pi=%d, %Ld-%Ld, bl=%d)\n",
306 file->f_file->f_dentry->d_inode->i_sb->s_id,
307 file->f_file->f_dentry->d_inode->i_ino,
308 lock->fl.fl_type, lock->fl.fl_pid,
309 (long long)lock->fl.fl_start,
310 (long long)lock->fl.fl_end,
311 wait);
314 /* Get existing block (in case client is busy-waiting) */
315 block = nlmsvc_lookup_block(file, lock, 0);
317 lock->fl.fl_flags |= FL_LOCKD;
319 again:
320 /* Lock file against concurrent access */
321 down(&file->f_sema);
323 if (!(conflock = posix_test_lock(file->f_file, &lock->fl))) {
324 error = posix_lock_file(file->f_file, &lock->fl);
326 if (block)
327 nlmsvc_delete_block(block, 0);
328 up(&file->f_sema);
330 dprintk("lockd: posix_lock_file returned %d\n", -error);
331 switch(-error) {
332 case 0:
333 return nlm_granted;
334 case EDEADLK:
335 return nlm_deadlock;
336 case EAGAIN:
337 return nlm_lck_denied;
338 default: /* includes ENOLCK */
339 return nlm_lck_denied_nolocks;
343 if (!wait) {
344 up(&file->f_sema);
345 return nlm_lck_denied;
348 if (posix_locks_deadlock(&lock->fl, conflock)) {
349 up(&file->f_sema);
350 return nlm_deadlock;
353 /* If we don't have a block, create and initialize it. Then
354 * retry because we may have slept in kmalloc. */
355 /* We have to release f_sema as nlmsvc_create_block may try to
356 * to claim it while doing host garbage collection */
357 if (block == NULL) {
358 up(&file->f_sema);
359 dprintk("lockd: blocking on this lock (allocating).\n");
360 if (!(block = nlmsvc_create_block(rqstp, file, lock, cookie)))
361 return nlm_lck_denied_nolocks;
362 goto again;
365 /* Append to list of blocked */
366 nlmsvc_insert_block(block, NLM_NEVER);
368 if (list_empty(&block->b_call.a_args.lock.fl.fl_block)) {
369 /* Now add block to block list of the conflicting lock
370 if we haven't done so. */
371 dprintk("lockd: blocking on this lock.\n");
372 posix_block_lock(conflock, &block->b_call.a_args.lock.fl);
375 up(&file->f_sema);
376 return nlm_lck_blocked;
380 * Test for presence of a conflicting lock.
383 nlmsvc_testlock(struct nlm_file *file, struct nlm_lock *lock,
384 struct nlm_lock *conflock)
386 struct file_lock *fl;
388 dprintk("lockd: nlmsvc_testlock(%s/%ld, ty=%d, %Ld-%Ld)\n",
389 file->f_file->f_dentry->d_inode->i_sb->s_id,
390 file->f_file->f_dentry->d_inode->i_ino,
391 lock->fl.fl_type,
392 (long long)lock->fl.fl_start,
393 (long long)lock->fl.fl_end);
395 if ((fl = posix_test_lock(file->f_file, &lock->fl)) != NULL) {
396 dprintk("lockd: conflicting lock(ty=%d, %Ld-%Ld)\n",
397 fl->fl_type, (long long)fl->fl_start,
398 (long long)fl->fl_end);
399 conflock->caller = "somehost"; /* FIXME */
400 conflock->oh.len = 0; /* don't return OH info */
401 conflock->fl = *fl;
402 return nlm_lck_denied;
405 return nlm_granted;
409 * Remove a lock.
410 * This implies a CANCEL call: We send a GRANT_MSG, the client replies
411 * with a GRANT_RES call which gets lost, and calls UNLOCK immediately
412 * afterwards. In this case the block will still be there, and hence
413 * must be removed.
416 nlmsvc_unlock(struct nlm_file *file, struct nlm_lock *lock)
418 int error;
420 dprintk("lockd: nlmsvc_unlock(%s/%ld, pi=%d, %Ld-%Ld)\n",
421 file->f_file->f_dentry->d_inode->i_sb->s_id,
422 file->f_file->f_dentry->d_inode->i_ino,
423 lock->fl.fl_pid,
424 (long long)lock->fl.fl_start,
425 (long long)lock->fl.fl_end);
427 /* First, cancel any lock that might be there */
428 nlmsvc_cancel_blocked(file, lock);
430 lock->fl.fl_type = F_UNLCK;
431 error = posix_lock_file(file->f_file, &lock->fl);
433 return (error < 0)? nlm_lck_denied_nolocks : nlm_granted;
437 * Cancel a previously blocked request.
439 * A cancel request always overrides any grant that may currently
440 * be in progress.
441 * The calling procedure must check whether the file can be closed.
444 nlmsvc_cancel_blocked(struct nlm_file *file, struct nlm_lock *lock)
446 struct nlm_block *block;
448 dprintk("lockd: nlmsvc_cancel(%s/%ld, pi=%d, %Ld-%Ld)\n",
449 file->f_file->f_dentry->d_inode->i_sb->s_id,
450 file->f_file->f_dentry->d_inode->i_ino,
451 lock->fl.fl_pid,
452 (long long)lock->fl.fl_start,
453 (long long)lock->fl.fl_end);
455 down(&file->f_sema);
456 if ((block = nlmsvc_lookup_block(file, lock, 1)) != NULL)
457 nlmsvc_delete_block(block, 1);
458 up(&file->f_sema);
459 return nlm_granted;
463 * Unblock a blocked lock request. This is a callback invoked from the
464 * VFS layer when a lock on which we blocked is removed.
466 * This function doesn't grant the blocked lock instantly, but rather moves
467 * the block to the head of nlm_blocked where it can be picked up by lockd.
469 static void
470 nlmsvc_notify_blocked(struct file_lock *fl)
472 struct nlm_block **bp, *block;
474 dprintk("lockd: VFS unblock notification for block %p\n", fl);
475 for (bp = &nlm_blocked; (block = *bp) != 0; bp = &block->b_next) {
476 if (nlm_compare_locks(&block->b_call.a_args.lock.fl, fl)) {
477 nlmsvc_insert_block(block, 0);
478 svc_wake_up(block->b_daemon);
479 return;
483 printk(KERN_WARNING "lockd: notification for unknown block!\n");
486 static int nlmsvc_same_owner(struct file_lock *fl1, struct file_lock *fl2)
488 return fl1->fl_owner == fl2->fl_owner && fl1->fl_pid == fl2->fl_pid;
491 struct lock_manager_operations nlmsvc_lock_operations = {
492 .fl_compare_owner = nlmsvc_same_owner,
493 .fl_notify = nlmsvc_notify_blocked,
497 * Try to claim a lock that was previously blocked.
499 * Note that we use both the RPC_GRANTED_MSG call _and_ an async
500 * RPC thread when notifying the client. This seems like overkill...
501 * Here's why:
502 * - we don't want to use a synchronous RPC thread, otherwise
503 * we might find ourselves hanging on a dead portmapper.
504 * - Some lockd implementations (e.g. HP) don't react to
505 * RPC_GRANTED calls; they seem to insist on RPC_GRANTED_MSG calls.
507 static void
508 nlmsvc_grant_blocked(struct nlm_block *block)
510 struct nlm_file *file = block->b_file;
511 struct nlm_lock *lock = &block->b_call.a_args.lock;
512 struct file_lock *conflock;
513 int error;
515 dprintk("lockd: grant blocked lock %p\n", block);
517 /* First thing is lock the file */
518 down(&file->f_sema);
520 /* Unlink block request from list */
521 nlmsvc_remove_block(block);
523 /* If b_granted is true this means we've been here before.
524 * Just retry the grant callback, possibly refreshing the RPC
525 * binding */
526 if (block->b_granted) {
527 nlm_rebind_host(block->b_host);
528 goto callback;
531 /* Try the lock operation again */
532 if ((conflock = posix_test_lock(file->f_file, &lock->fl)) != NULL) {
533 /* Bummer, we blocked again */
534 dprintk("lockd: lock still blocked\n");
535 nlmsvc_insert_block(block, NLM_NEVER);
536 posix_block_lock(conflock, &lock->fl);
537 up(&file->f_sema);
538 return;
541 /* Alright, no conflicting lock. Now lock it for real. If the
542 * following yields an error, this is most probably due to low
543 * memory. Retry the lock in a few seconds.
545 if ((error = posix_lock_file(file->f_file, &lock->fl)) < 0) {
546 printk(KERN_WARNING "lockd: unexpected error %d in %s!\n",
547 -error, __FUNCTION__);
548 nlmsvc_insert_block(block, 10 * HZ);
549 up(&file->f_sema);
550 return;
553 callback:
554 /* Lock was granted by VFS. */
555 dprintk("lockd: GRANTing blocked lock.\n");
556 block->b_granted = 1;
557 block->b_incall = 1;
559 /* Schedule next grant callback in 30 seconds */
560 nlmsvc_insert_block(block, 30 * HZ);
562 /* Call the client */
563 nlm_get_host(block->b_call.a_host);
564 if (nlmsvc_async_call(&block->b_call, NLMPROC_GRANTED_MSG,
565 nlmsvc_grant_callback) < 0)
566 nlm_release_host(block->b_call.a_host);
567 up(&file->f_sema);
571 * This is the callback from the RPC layer when the NLM_GRANTED_MSG
572 * RPC call has succeeded or timed out.
573 * Like all RPC callbacks, it is invoked by the rpciod process, so it
574 * better not sleep. Therefore, we put the blocked lock on the nlm_blocked
575 * chain once more in order to have it removed by lockd itself (which can
576 * then sleep on the file semaphore without disrupting e.g. the nfs client).
578 static void
579 nlmsvc_grant_callback(struct rpc_task *task)
581 struct nlm_rqst *call = (struct nlm_rqst *) task->tk_calldata;
582 struct nlm_block *block;
583 unsigned long timeout;
584 struct sockaddr_in *peer_addr = RPC_PEERADDR(task->tk_client);
586 dprintk("lockd: GRANT_MSG RPC callback\n");
587 dprintk("callback: looking for cookie %x, host (%08x)\n",
588 *(unsigned int *)(call->a_args.cookie.data),
589 ntohl(peer_addr->sin_addr.s_addr));
590 if (!(block = nlmsvc_find_block(&call->a_args.cookie, peer_addr))) {
591 dprintk("lockd: no block for cookie %x, host (%08x)\n",
592 *(u32 *)(call->a_args.cookie.data),
593 ntohl(peer_addr->sin_addr.s_addr));
594 return;
597 /* Technically, we should down the file semaphore here. Since we
598 * move the block towards the head of the queue only, no harm
599 * can be done, though. */
600 if (task->tk_status < 0) {
601 /* RPC error: Re-insert for retransmission */
602 timeout = 10 * HZ;
603 } else if (block->b_done) {
604 /* Block already removed, kill it for real */
605 timeout = 0;
606 } else {
607 /* Call was successful, now wait for client callback */
608 timeout = 60 * HZ;
610 nlmsvc_insert_block(block, timeout);
611 svc_wake_up(block->b_daemon);
612 block->b_incall = 0;
614 nlm_release_host(call->a_host);
618 * We received a GRANT_RES callback. Try to find the corresponding
619 * block.
621 void
622 nlmsvc_grant_reply(struct svc_rqst *rqstp, struct nlm_cookie *cookie, u32 status)
624 struct nlm_block *block;
625 struct nlm_file *file;
627 dprintk("grant_reply: looking for cookie %x, host (%08x), s=%d \n",
628 *(unsigned int *)(cookie->data),
629 ntohl(rqstp->rq_addr.sin_addr.s_addr), status);
630 if (!(block = nlmsvc_find_block(cookie, &rqstp->rq_addr)))
631 return;
632 file = block->b_file;
634 file->f_count++;
635 down(&file->f_sema);
636 if ((block = nlmsvc_find_block(cookie,&rqstp->rq_addr)) != NULL) {
637 if (status == NLM_LCK_DENIED_GRACE_PERIOD) {
638 /* Try again in a couple of seconds */
639 nlmsvc_insert_block(block, 10 * HZ);
640 block = NULL;
641 } else {
642 /* Lock is now held by client, or has been rejected.
643 * In both cases, the block should be removed. */
644 file->f_count++;
645 up(&file->f_sema);
646 if (status == NLM_LCK_GRANTED)
647 nlmsvc_delete_block(block, 0);
648 else
649 nlmsvc_delete_block(block, 1);
652 if (!block)
653 up(&file->f_sema);
654 nlm_release_file(file);
658 * Retry all blocked locks that have been notified. This is where lockd
659 * picks up locks that can be granted, or grant notifications that must
660 * be retransmitted.
662 unsigned long
663 nlmsvc_retry_blocked(void)
665 struct nlm_block *block;
667 dprintk("nlmsvc_retry_blocked(%p, when=%ld)\n",
668 nlm_blocked,
669 nlm_blocked? nlm_blocked->b_when : 0);
670 while ((block = nlm_blocked) != 0) {
671 if (block->b_when == NLM_NEVER)
672 break;
673 if (time_after(block->b_when,jiffies))
674 break;
675 dprintk("nlmsvc_retry_blocked(%p, when=%ld, done=%d)\n",
676 block, block->b_when, block->b_done);
677 if (block->b_done)
678 nlmsvc_delete_block(block, 0);
679 else
680 nlmsvc_grant_blocked(block);
683 if ((block = nlm_blocked) && block->b_when != NLM_NEVER)
684 return (block->b_when - jiffies);
686 return MAX_SCHEDULE_TIMEOUT;