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
39 #define nlm_deadlock nlm_lck_denied
42 static void nlmsvc_release_block(struct nlm_block *block);
43 static void nlmsvc_insert_block(struct nlm_block *block, unsigned long);
44 static int nlmsvc_remove_block(struct nlm_block *block);
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;
51 * The list of blocked locks to retry
53 static struct nlm_block * nlm_blocked;
56 * Insert a blocked lock into the global list
59 nlmsvc_insert_block(struct nlm_block *block, unsigned long when)
61 struct nlm_block **bp, *b;
63 dprintk("lockd: nlmsvc_insert_block(%p, %ld)\n", block, when);
64 kref_get(&block->b_count);
66 nlmsvc_remove_block(block);
68 if (when != NLM_NEVER) {
69 if ((when += jiffies) == NLM_NEVER)
71 while ((b = *bp) && time_before_eq(b->b_when,when) && b->b_when != NLM_NEVER)
74 while ((b = *bp) != 0)
84 * Remove a block from the global list
87 nlmsvc_remove_block(struct nlm_block *block)
89 struct nlm_block **bp, *b;
93 for (bp = &nlm_blocked; (b = *bp) != 0; bp = &b->b_next) {
97 nlmsvc_release_block(block);
106 * Find a block for a given lock
108 static struct nlm_block *
109 nlmsvc_lookup_block(struct nlm_file *file, struct nlm_lock *lock)
111 struct nlm_block **head, *block;
112 struct file_lock *fl;
114 dprintk("lockd: nlmsvc_lookup_block f=%p pd=%d %Ld-%Ld ty=%d\n",
115 file, lock->fl.fl_pid,
116 (long long)lock->fl.fl_start,
117 (long long)lock->fl.fl_end, lock->fl.fl_type);
118 for (head = &nlm_blocked; (block = *head) != 0; head = &block->b_next) {
119 fl = &block->b_call->a_args.lock.fl;
120 dprintk("lockd: check f=%p pd=%d %Ld-%Ld ty=%d cookie=%s\n",
121 block->b_file, fl->fl_pid,
122 (long long)fl->fl_start,
123 (long long)fl->fl_end, fl->fl_type,
124 nlmdbg_cookie2a(&block->b_call->a_args.cookie));
125 if (block->b_file == file && nlm_compare_locks(fl, &lock->fl)) {
126 kref_get(&block->b_count);
134 static inline int nlm_cookie_match(struct nlm_cookie *a, struct nlm_cookie *b)
138 if(memcmp(a->data,b->data,a->len))
144 * Find a block with a given NLM cookie.
146 static inline struct nlm_block *
147 nlmsvc_find_block(struct nlm_cookie *cookie, struct sockaddr_in *sin)
149 struct nlm_block *block;
151 for (block = nlm_blocked; block; block = block->b_next) {
152 dprintk("cookie: head of blocked queue %p, block %p\n",
154 if (nlm_cookie_match(&block->b_call->a_args.cookie,cookie)
155 && nlm_cmp_addr(sin, &block->b_host->h_addr))
160 kref_get(&block->b_count);
165 * Create a block and initialize it.
167 * Note: we explicitly set the cookie of the grant reply to that of
168 * the blocked lock request. The spec explicitly mentions that the client
169 * should _not_ rely on the callback containing the same cookie as the
170 * request, but (as I found out later) that's because some implementations
171 * do just this. Never mind the standards comittees, they support our
172 * logging industries.
174 static inline struct nlm_block *
175 nlmsvc_create_block(struct svc_rqst *rqstp, struct nlm_file *file,
176 struct nlm_lock *lock, struct nlm_cookie *cookie)
178 struct nlm_block *block;
179 struct nlm_host *host;
180 struct nlm_rqst *call = NULL;
182 /* Create host handle for callback */
183 host = nlmsvc_lookup_host(rqstp);
187 call = nlm_alloc_call(host);
191 /* Allocate memory for block, and initialize arguments */
192 block = kzalloc(sizeof(*block), GFP_KERNEL);
195 kref_init(&block->b_count);
197 if (!nlmsvc_setgrantargs(call, lock))
200 /* Set notifier function for VFS, and init args */
201 call->a_args.lock.fl.fl_flags |= FL_SLEEP;
202 call->a_args.lock.fl.fl_lmops = &nlmsvc_lock_operations;
203 call->a_args.cookie = *cookie; /* see above */
205 dprintk("lockd: created block %p...\n", block);
207 /* Create and initialize the block */
208 block->b_daemon = rqstp->rq_server;
209 block->b_host = host;
210 block->b_file = file;
213 /* Add to file's list of blocks */
214 block->b_fnext = file->f_blocks;
215 file->f_blocks = block;
217 /* Set up RPC arguments for callback */
218 block->b_call = call;
219 call->a_flags = RPC_TASK_ASYNC;
220 call->a_block = block;
227 nlm_release_call(call);
232 * Delete a block. If the lock was cancelled or the grant callback
233 * failed, unlock is set to 1.
234 * It is the caller's responsibility to check whether the file
235 * can be closed hereafter.
237 static int nlmsvc_unlink_block(struct nlm_block *block)
240 dprintk("lockd: unlinking block %p...\n", block);
242 /* Remove block from list */
243 status = posix_unblock_lock(block->b_file->f_file, &block->b_call->a_args.lock.fl);
244 nlmsvc_remove_block(block);
248 static void nlmsvc_free_block(struct kref *kref)
250 struct nlm_block *block = container_of(kref, struct nlm_block, b_count);
251 struct nlm_file *file = block->b_file;
252 struct nlm_block **bp;
254 dprintk("lockd: freeing block %p...\n", block);
257 /* Remove block from file's list of blocks */
258 for (bp = &file->f_blocks; *bp; bp = &(*bp)->b_fnext) {
260 *bp = block->b_fnext;
266 nlmsvc_freegrantargs(block->b_call);
267 nlm_release_call(block->b_call);
268 nlm_release_file(block->b_file);
272 static void nlmsvc_release_block(struct nlm_block *block)
275 kref_put(&block->b_count, nlmsvc_free_block);
278 static void nlmsvc_act_mark(struct nlm_host *host, struct nlm_file *file)
280 struct nlm_block *block;
283 for (block = file->f_blocks; block != NULL; block = block->b_fnext)
284 block->b_host->h_inuse = 1;
288 static void nlmsvc_act_unlock(struct nlm_host *host, struct nlm_file *file)
290 struct nlm_block *block;
294 for (block = file->f_blocks; block != NULL; block = block->b_fnext) {
295 if (host != NULL && host != block->b_host)
297 if (!block->b_queued)
299 kref_get(&block->b_count);
301 nlmsvc_unlink_block(block);
302 nlmsvc_release_block(block);
309 * Loop over all blocks and perform the action specified.
310 * (NLM_ACT_CHECK handled by nlmsvc_inspect_file).
313 nlmsvc_traverse_blocks(struct nlm_host *host, struct nlm_file *file, int action)
315 if (action == NLM_ACT_MARK)
316 nlmsvc_act_mark(host, file);
318 nlmsvc_act_unlock(host, file);
323 * Initialize arguments for GRANTED call. The nlm_rqst structure
324 * has been cleared already.
326 static int nlmsvc_setgrantargs(struct nlm_rqst *call, struct nlm_lock *lock)
328 locks_copy_lock(&call->a_args.lock.fl, &lock->fl);
329 memcpy(&call->a_args.lock.fh, &lock->fh, sizeof(call->a_args.lock.fh));
330 call->a_args.lock.caller = system_utsname.nodename;
331 call->a_args.lock.oh.len = lock->oh.len;
333 /* set default data area */
334 call->a_args.lock.oh.data = call->a_owner;
335 call->a_args.lock.svid = lock->fl.fl_pid;
337 if (lock->oh.len > NLMCLNT_OHSIZE) {
338 void *data = kmalloc(lock->oh.len, GFP_KERNEL);
341 call->a_args.lock.oh.data = (u8 *) data;
344 memcpy(call->a_args.lock.oh.data, lock->oh.data, lock->oh.len);
348 static void nlmsvc_freegrantargs(struct nlm_rqst *call)
350 if (call->a_args.lock.oh.data != call->a_owner)
351 kfree(call->a_args.lock.oh.data);
355 * Attempt to establish a lock, and if it can't be granted, block it
359 nlmsvc_lock(struct svc_rqst *rqstp, struct nlm_file *file,
360 struct nlm_lock *lock, int wait, struct nlm_cookie *cookie)
362 struct nlm_block *block, *newblock = NULL;
366 dprintk("lockd: nlmsvc_lock(%s/%ld, ty=%d, pi=%d, %Ld-%Ld, bl=%d)\n",
367 file->f_file->f_dentry->d_inode->i_sb->s_id,
368 file->f_file->f_dentry->d_inode->i_ino,
369 lock->fl.fl_type, lock->fl.fl_pid,
370 (long long)lock->fl.fl_start,
371 (long long)lock->fl.fl_end,
375 lock->fl.fl_flags &= ~FL_SLEEP;
377 /* Lock file against concurrent access */
379 /* Get existing block (in case client is busy-waiting) */
380 block = nlmsvc_lookup_block(file, lock);
382 if (newblock != NULL)
383 lock = &newblock->b_call->a_args.lock;
385 lock = &block->b_call->a_args.lock;
387 error = posix_lock_file(file->f_file, &lock->fl);
388 lock->fl.fl_flags &= ~FL_SLEEP;
390 dprintk("lockd: posix_lock_file returned %d\n", error);
401 default: /* includes ENOLCK */
402 ret = nlm_lck_denied_nolocks;
406 ret = nlm_lck_denied;
410 ret = nlm_lck_blocked;
414 /* If we don't have a block, create and initialize it. Then
415 * retry because we may have slept in kmalloc. */
416 /* We have to release f_sema as nlmsvc_create_block may try to
417 * to claim it while doing host garbage collection */
418 if (newblock == NULL) {
420 dprintk("lockd: blocking on this lock (allocating).\n");
421 if (!(newblock = nlmsvc_create_block(rqstp, file, lock, cookie)))
422 return nlm_lck_denied_nolocks;
426 /* Append to list of blocked */
427 nlmsvc_insert_block(newblock, NLM_NEVER);
430 nlmsvc_release_block(newblock);
431 nlmsvc_release_block(block);
432 dprintk("lockd: nlmsvc_lock returned %u\n", ret);
437 * Test for presence of a conflicting lock.
440 nlmsvc_testlock(struct nlm_file *file, struct nlm_lock *lock,
441 struct nlm_lock *conflock)
443 dprintk("lockd: nlmsvc_testlock(%s/%ld, ty=%d, %Ld-%Ld)\n",
444 file->f_file->f_dentry->d_inode->i_sb->s_id,
445 file->f_file->f_dentry->d_inode->i_ino,
447 (long long)lock->fl.fl_start,
448 (long long)lock->fl.fl_end);
450 if (posix_test_lock(file->f_file, &lock->fl, &conflock->fl)) {
451 dprintk("lockd: conflicting lock(ty=%d, %Ld-%Ld)\n",
452 conflock->fl.fl_type,
453 (long long)conflock->fl.fl_start,
454 (long long)conflock->fl.fl_end);
455 conflock->caller = "somehost"; /* FIXME */
456 conflock->oh.len = 0; /* don't return OH info */
457 conflock->svid = conflock->fl.fl_pid;
458 return nlm_lck_denied;
466 * This implies a CANCEL call: We send a GRANT_MSG, the client replies
467 * with a GRANT_RES call which gets lost, and calls UNLOCK immediately
468 * afterwards. In this case the block will still be there, and hence
472 nlmsvc_unlock(struct nlm_file *file, struct nlm_lock *lock)
476 dprintk("lockd: nlmsvc_unlock(%s/%ld, pi=%d, %Ld-%Ld)\n",
477 file->f_file->f_dentry->d_inode->i_sb->s_id,
478 file->f_file->f_dentry->d_inode->i_ino,
480 (long long)lock->fl.fl_start,
481 (long long)lock->fl.fl_end);
483 /* First, cancel any lock that might be there */
484 nlmsvc_cancel_blocked(file, lock);
486 lock->fl.fl_type = F_UNLCK;
487 error = posix_lock_file(file->f_file, &lock->fl);
489 return (error < 0)? nlm_lck_denied_nolocks : nlm_granted;
493 * Cancel a previously blocked request.
495 * A cancel request always overrides any grant that may currently
497 * The calling procedure must check whether the file can be closed.
500 nlmsvc_cancel_blocked(struct nlm_file *file, struct nlm_lock *lock)
502 struct nlm_block *block;
505 dprintk("lockd: nlmsvc_cancel(%s/%ld, pi=%d, %Ld-%Ld)\n",
506 file->f_file->f_dentry->d_inode->i_sb->s_id,
507 file->f_file->f_dentry->d_inode->i_ino,
509 (long long)lock->fl.fl_start,
510 (long long)lock->fl.fl_end);
513 block = nlmsvc_lookup_block(file, lock);
516 status = nlmsvc_unlink_block(block);
517 nlmsvc_release_block(block);
519 return status ? nlm_lck_denied : nlm_granted;
523 * Unblock a blocked lock request. This is a callback invoked from the
524 * VFS layer when a lock on which we blocked is removed.
526 * This function doesn't grant the blocked lock instantly, but rather moves
527 * the block to the head of nlm_blocked where it can be picked up by lockd.
530 nlmsvc_notify_blocked(struct file_lock *fl)
532 struct nlm_block **bp, *block;
534 dprintk("lockd: VFS unblock notification for block %p\n", fl);
535 for (bp = &nlm_blocked; (block = *bp) != 0; bp = &block->b_next) {
536 if (nlm_compare_locks(&block->b_call->a_args.lock.fl, fl)) {
537 nlmsvc_insert_block(block, 0);
538 svc_wake_up(block->b_daemon);
543 printk(KERN_WARNING "lockd: notification for unknown block!\n");
546 static int nlmsvc_same_owner(struct file_lock *fl1, struct file_lock *fl2)
548 return fl1->fl_owner == fl2->fl_owner && fl1->fl_pid == fl2->fl_pid;
551 struct lock_manager_operations nlmsvc_lock_operations = {
552 .fl_compare_owner = nlmsvc_same_owner,
553 .fl_notify = nlmsvc_notify_blocked,
557 * Try to claim a lock that was previously blocked.
559 * Note that we use both the RPC_GRANTED_MSG call _and_ an async
560 * RPC thread when notifying the client. This seems like overkill...
562 * - we don't want to use a synchronous RPC thread, otherwise
563 * we might find ourselves hanging on a dead portmapper.
564 * - Some lockd implementations (e.g. HP) don't react to
565 * RPC_GRANTED calls; they seem to insist on RPC_GRANTED_MSG calls.
568 nlmsvc_grant_blocked(struct nlm_block *block)
570 struct nlm_file *file = block->b_file;
571 struct nlm_lock *lock = &block->b_call->a_args.lock;
574 dprintk("lockd: grant blocked lock %p\n", block);
576 /* Unlink block request from list */
577 nlmsvc_unlink_block(block);
579 /* If b_granted is true this means we've been here before.
580 * Just retry the grant callback, possibly refreshing the RPC
582 if (block->b_granted) {
583 nlm_rebind_host(block->b_host);
587 /* Try the lock operation again */
588 lock->fl.fl_flags |= FL_SLEEP;
589 error = posix_lock_file(file->f_file, &lock->fl);
590 lock->fl.fl_flags &= ~FL_SLEEP;
596 dprintk("lockd: lock still blocked\n");
597 nlmsvc_insert_block(block, NLM_NEVER);
600 printk(KERN_WARNING "lockd: unexpected error %d in %s!\n",
601 -error, __FUNCTION__);
602 nlmsvc_insert_block(block, 10 * HZ);
607 /* Lock was granted by VFS. */
608 dprintk("lockd: GRANTing blocked lock.\n");
609 block->b_granted = 1;
611 /* Schedule next grant callback in 30 seconds */
612 nlmsvc_insert_block(block, 30 * HZ);
614 /* Call the client */
615 kref_get(&block->b_count);
616 if (nlm_async_call(block->b_call, NLMPROC_GRANTED_MSG,
617 &nlmsvc_grant_ops) < 0)
618 nlmsvc_release_block(block);
622 * This is the callback from the RPC layer when the NLM_GRANTED_MSG
623 * RPC call has succeeded or timed out.
624 * Like all RPC callbacks, it is invoked by the rpciod process, so it
625 * better not sleep. Therefore, we put the blocked lock on the nlm_blocked
626 * chain once more in order to have it removed by lockd itself (which can
627 * then sleep on the file semaphore without disrupting e.g. the nfs client).
629 static void nlmsvc_grant_callback(struct rpc_task *task, void *data)
631 struct nlm_rqst *call = data;
632 struct nlm_block *block = call->a_block;
633 unsigned long timeout;
635 dprintk("lockd: GRANT_MSG RPC callback\n");
637 /* Technically, we should down the file semaphore here. Since we
638 * move the block towards the head of the queue only, no harm
639 * can be done, though. */
640 if (task->tk_status < 0) {
641 /* RPC error: Re-insert for retransmission */
643 } else if (block->b_done) {
644 /* Block already removed, kill it for real */
647 /* Call was successful, now wait for client callback */
650 nlmsvc_insert_block(block, timeout);
651 svc_wake_up(block->b_daemon);
654 void nlmsvc_grant_release(void *data)
656 struct nlm_rqst *call = data;
658 nlmsvc_release_block(call->a_block);
661 static const struct rpc_call_ops nlmsvc_grant_ops = {
662 .rpc_call_done = nlmsvc_grant_callback,
663 .rpc_release = nlmsvc_grant_release,
667 * We received a GRANT_RES callback. Try to find the corresponding
671 nlmsvc_grant_reply(struct svc_rqst *rqstp, struct nlm_cookie *cookie, u32 status)
673 struct nlm_block *block;
674 struct nlm_file *file;
676 dprintk("grant_reply: looking for cookie %x, host (%08x), s=%d \n",
677 *(unsigned int *)(cookie->data),
678 ntohl(rqstp->rq_addr.sin_addr.s_addr), status);
679 if (!(block = nlmsvc_find_block(cookie, &rqstp->rq_addr)))
681 file = block->b_file;
684 if (status == NLM_LCK_DENIED_GRACE_PERIOD) {
685 /* Try again in a couple of seconds */
686 nlmsvc_insert_block(block, 10 * HZ);
688 /* Lock is now held by client, or has been rejected.
689 * In both cases, the block should be removed. */
690 nlmsvc_unlink_block(block);
693 nlmsvc_release_block(block);
697 * Retry all blocked locks that have been notified. This is where lockd
698 * picks up locks that can be granted, or grant notifications that must
702 nlmsvc_retry_blocked(void)
704 struct nlm_block *block;
706 dprintk("nlmsvc_retry_blocked(%p, when=%ld)\n",
708 nlm_blocked? nlm_blocked->b_when : 0);
709 while ((block = nlm_blocked) != 0) {
710 if (block->b_when == NLM_NEVER)
712 if (time_after(block->b_when,jiffies))
714 dprintk("nlmsvc_retry_blocked(%p, when=%ld, done=%d)\n",
715 block, block->b_when, block->b_done);
716 kref_get(&block->b_count);
718 nlmsvc_unlink_block(block);
720 nlmsvc_grant_blocked(block);
721 nlmsvc_release_block(block);
724 if ((block = nlm_blocked) && block->b_when != NLM_NEVER)
725 return (block->b_when - jiffies);
727 return MAX_SCHEDULE_TIMEOUT;