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/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>
33 #define NLMDBG_FACILITY NLMDBG_SVCLOCK
35 #ifdef CONFIG_LOCKD_V4
36 #define nlm_deadlock nlm4_deadlock
38 #define nlm_deadlock nlm_lck_denied
41 static void nlmsvc_release_block(struct nlm_block *block);
42 static void nlmsvc_insert_block(struct nlm_block *block, unsigned long);
43 static void nlmsvc_remove_block(struct nlm_block *block);
45 static int nlmsvc_setgrantargs(struct nlm_rqst *call, struct nlm_lock *lock);
46 static void nlmsvc_freegrantargs(struct nlm_rqst *call);
47 static const struct rpc_call_ops nlmsvc_grant_ops;
50 * The list of blocked locks to retry
52 static LIST_HEAD(nlm_blocked);
55 * Insert a blocked lock into the global list
58 nlmsvc_insert_block(struct nlm_block *block, unsigned long when)
61 struct list_head *pos;
63 dprintk("lockd: nlmsvc_insert_block(%p, %ld)\n", block, when);
64 if (list_empty(&block->b_list)) {
65 kref_get(&block->b_count);
67 list_del_init(&block->b_list);
71 if (when != NLM_NEVER) {
72 if ((when += jiffies) == NLM_NEVER)
74 list_for_each(pos, &nlm_blocked) {
75 b = list_entry(pos, struct nlm_block, b_list);
76 if (time_after(b->b_when,when) || b->b_when == NLM_NEVER)
79 /* On normal exit from the loop, pos == &nlm_blocked,
80 * so we will be adding to the end of the list - good
84 list_add_tail(&block->b_list, pos);
89 * Remove a block from the global list
92 nlmsvc_remove_block(struct nlm_block *block)
94 if (!list_empty(&block->b_list)) {
95 list_del_init(&block->b_list);
96 nlmsvc_release_block(block);
101 * Find a block for a given lock
103 static struct nlm_block *
104 nlmsvc_lookup_block(struct nlm_file *file, struct nlm_lock *lock)
106 struct nlm_block *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 list_for_each_entry(block, &nlm_blocked, b_list) {
114 fl = &block->b_call->a_args.lock.fl;
115 dprintk("lockd: check f=%p pd=%d %Ld-%Ld ty=%d cookie=%s\n",
116 block->b_file, fl->fl_pid,
117 (long long)fl->fl_start,
118 (long long)fl->fl_end, fl->fl_type,
119 nlmdbg_cookie2a(&block->b_call->a_args.cookie));
120 if (block->b_file == file && nlm_compare_locks(fl, &lock->fl)) {
121 kref_get(&block->b_count);
129 static inline int nlm_cookie_match(struct nlm_cookie *a, struct nlm_cookie *b)
133 if(memcmp(a->data,b->data,a->len))
139 * Find a block with a given NLM cookie.
141 static inline struct nlm_block *
142 nlmsvc_find_block(struct nlm_cookie *cookie, struct sockaddr_in *sin)
144 struct nlm_block *block;
146 list_for_each_entry(block, &nlm_blocked, b_list) {
147 if (nlm_cookie_match(&block->b_call->a_args.cookie,cookie)
148 && nlm_cmp_addr(sin, &block->b_host->h_addr))
155 kref_get(&block->b_count);
160 * Create a block and initialize it.
162 * Note: we explicitly set the cookie of the grant reply to that of
163 * the blocked lock request. The spec explicitly mentions that the client
164 * should _not_ rely on the callback containing the same cookie as the
165 * request, but (as I found out later) that's because some implementations
166 * do just this. Never mind the standards comittees, they support our
167 * logging industries.
169 static inline struct nlm_block *
170 nlmsvc_create_block(struct svc_rqst *rqstp, struct nlm_file *file,
171 struct nlm_lock *lock, struct nlm_cookie *cookie)
173 struct nlm_block *block;
174 struct nlm_host *host;
175 struct nlm_rqst *call = NULL;
177 /* Create host handle for callback */
178 host = nlmsvc_lookup_host(rqstp, lock->caller, lock->len);
182 call = nlm_alloc_call(host);
186 /* Allocate memory for block, and initialize arguments */
187 block = kzalloc(sizeof(*block), GFP_KERNEL);
190 kref_init(&block->b_count);
191 INIT_LIST_HEAD(&block->b_list);
192 INIT_LIST_HEAD(&block->b_flist);
194 if (!nlmsvc_setgrantargs(call, lock))
197 /* Set notifier function for VFS, and init args */
198 call->a_args.lock.fl.fl_flags |= FL_SLEEP;
199 call->a_args.lock.fl.fl_lmops = &nlmsvc_lock_operations;
200 call->a_args.cookie = *cookie; /* see above */
202 dprintk("lockd: created block %p...\n", block);
204 /* Create and initialize the block */
205 block->b_daemon = rqstp->rq_server;
206 block->b_host = host;
207 block->b_file = file;
210 /* Add to file's list of blocks */
211 list_add(&block->b_flist, &file->f_blocks);
213 /* Set up RPC arguments for callback */
214 block->b_call = call;
215 call->a_flags = RPC_TASK_ASYNC;
216 call->a_block = block;
223 nlm_release_call(call);
228 * Delete a block. If the lock was cancelled or the grant callback
229 * failed, unlock is set to 1.
230 * It is the caller's responsibility to check whether the file
231 * can be closed hereafter.
233 static int nlmsvc_unlink_block(struct nlm_block *block)
236 dprintk("lockd: unlinking block %p...\n", block);
238 /* Remove block from list */
239 status = posix_unblock_lock(block->b_file->f_file, &block->b_call->a_args.lock.fl);
240 nlmsvc_remove_block(block);
244 static void nlmsvc_free_block(struct kref *kref)
246 struct nlm_block *block = container_of(kref, struct nlm_block, b_count);
247 struct nlm_file *file = block->b_file;
249 dprintk("lockd: freeing block %p...\n", block);
251 /* Remove block from file's list of blocks */
253 list_del_init(&block->b_flist);
256 nlmsvc_freegrantargs(block->b_call);
257 nlm_release_call(block->b_call);
258 nlm_release_file(block->b_file);
262 static void nlmsvc_release_block(struct nlm_block *block)
265 kref_put(&block->b_count, nlmsvc_free_block);
268 static void nlmsvc_act_mark(struct nlm_host *host, struct nlm_file *file)
270 struct nlm_block *block;
273 list_for_each_entry(block, &file->f_blocks, b_flist)
274 block->b_host->h_inuse = 1;
278 static void nlmsvc_act_unlock(struct nlm_host *host, struct nlm_file *file)
280 struct nlm_block *block, *next;
284 list_for_each_entry_safe(block, next, &file->f_blocks, b_flist) {
285 if (host != NULL && host != block->b_host)
287 /* Do not destroy blocks that are not on
288 * the global retry list - why? */
289 if (list_empty(&block->b_list))
291 kref_get(&block->b_count);
293 nlmsvc_unlink_block(block);
294 nlmsvc_release_block(block);
301 * Loop over all blocks and perform the action specified.
302 * (NLM_ACT_CHECK handled by nlmsvc_inspect_file).
305 nlmsvc_traverse_blocks(struct nlm_host *host, struct nlm_file *file, int action)
307 if (action == NLM_ACT_MARK)
308 nlmsvc_act_mark(host, file);
310 nlmsvc_act_unlock(host, file);
314 * Initialize arguments for GRANTED call. The nlm_rqst structure
315 * has been cleared already.
317 static int nlmsvc_setgrantargs(struct nlm_rqst *call, struct nlm_lock *lock)
319 locks_copy_lock(&call->a_args.lock.fl, &lock->fl);
320 memcpy(&call->a_args.lock.fh, &lock->fh, sizeof(call->a_args.lock.fh));
321 call->a_args.lock.caller = utsname()->nodename;
322 call->a_args.lock.oh.len = lock->oh.len;
324 /* set default data area */
325 call->a_args.lock.oh.data = call->a_owner;
326 call->a_args.lock.svid = lock->fl.fl_pid;
328 if (lock->oh.len > NLMCLNT_OHSIZE) {
329 void *data = kmalloc(lock->oh.len, GFP_KERNEL);
332 call->a_args.lock.oh.data = (u8 *) data;
335 memcpy(call->a_args.lock.oh.data, lock->oh.data, lock->oh.len);
339 static void nlmsvc_freegrantargs(struct nlm_rqst *call)
341 if (call->a_args.lock.oh.data != call->a_owner)
342 kfree(call->a_args.lock.oh.data);
346 * Attempt to establish a lock, and if it can't be granted, block it
350 nlmsvc_lock(struct svc_rqst *rqstp, struct nlm_file *file,
351 struct nlm_lock *lock, int wait, struct nlm_cookie *cookie)
353 struct nlm_block *block, *newblock = NULL;
357 dprintk("lockd: nlmsvc_lock(%s/%ld, ty=%d, pi=%d, %Ld-%Ld, bl=%d)\n",
358 file->f_file->f_dentry->d_inode->i_sb->s_id,
359 file->f_file->f_dentry->d_inode->i_ino,
360 lock->fl.fl_type, lock->fl.fl_pid,
361 (long long)lock->fl.fl_start,
362 (long long)lock->fl.fl_end,
366 lock->fl.fl_flags &= ~FL_SLEEP;
368 /* Lock file against concurrent access */
370 /* Get existing block (in case client is busy-waiting) */
371 block = nlmsvc_lookup_block(file, lock);
373 if (newblock != NULL)
374 lock = &newblock->b_call->a_args.lock;
376 lock = &block->b_call->a_args.lock;
378 error = posix_lock_file(file->f_file, &lock->fl);
379 lock->fl.fl_flags &= ~FL_SLEEP;
381 dprintk("lockd: posix_lock_file returned %d\n", error);
392 default: /* includes ENOLCK */
393 ret = nlm_lck_denied_nolocks;
397 ret = nlm_lck_denied;
401 ret = nlm_lck_blocked;
405 /* If we don't have a block, create and initialize it. Then
406 * retry because we may have slept in kmalloc. */
407 /* We have to release f_sema as nlmsvc_create_block may try to
408 * to claim it while doing host garbage collection */
409 if (newblock == NULL) {
411 dprintk("lockd: blocking on this lock (allocating).\n");
412 if (!(newblock = nlmsvc_create_block(rqstp, file, lock, cookie)))
413 return nlm_lck_denied_nolocks;
417 /* Append to list of blocked */
418 nlmsvc_insert_block(newblock, NLM_NEVER);
421 nlmsvc_release_block(newblock);
422 nlmsvc_release_block(block);
423 dprintk("lockd: nlmsvc_lock returned %u\n", ret);
428 * Test for presence of a conflicting lock.
431 nlmsvc_testlock(struct nlm_file *file, struct nlm_lock *lock,
432 struct nlm_lock *conflock)
434 dprintk("lockd: nlmsvc_testlock(%s/%ld, ty=%d, %Ld-%Ld)\n",
435 file->f_file->f_dentry->d_inode->i_sb->s_id,
436 file->f_file->f_dentry->d_inode->i_ino,
438 (long long)lock->fl.fl_start,
439 (long long)lock->fl.fl_end);
441 if (posix_test_lock(file->f_file, &lock->fl, &conflock->fl)) {
442 dprintk("lockd: conflicting lock(ty=%d, %Ld-%Ld)\n",
443 conflock->fl.fl_type,
444 (long long)conflock->fl.fl_start,
445 (long long)conflock->fl.fl_end);
446 conflock->caller = "somehost"; /* FIXME */
447 conflock->len = strlen(conflock->caller);
448 conflock->oh.len = 0; /* don't return OH info */
449 conflock->svid = conflock->fl.fl_pid;
450 return nlm_lck_denied;
458 * This implies a CANCEL call: We send a GRANT_MSG, the client replies
459 * with a GRANT_RES call which gets lost, and calls UNLOCK immediately
460 * afterwards. In this case the block will still be there, and hence
464 nlmsvc_unlock(struct nlm_file *file, struct nlm_lock *lock)
468 dprintk("lockd: nlmsvc_unlock(%s/%ld, pi=%d, %Ld-%Ld)\n",
469 file->f_file->f_dentry->d_inode->i_sb->s_id,
470 file->f_file->f_dentry->d_inode->i_ino,
472 (long long)lock->fl.fl_start,
473 (long long)lock->fl.fl_end);
475 /* First, cancel any lock that might be there */
476 nlmsvc_cancel_blocked(file, lock);
478 lock->fl.fl_type = F_UNLCK;
479 error = posix_lock_file(file->f_file, &lock->fl);
481 return (error < 0)? nlm_lck_denied_nolocks : nlm_granted;
485 * Cancel a previously blocked request.
487 * A cancel request always overrides any grant that may currently
489 * The calling procedure must check whether the file can be closed.
492 nlmsvc_cancel_blocked(struct nlm_file *file, struct nlm_lock *lock)
494 struct nlm_block *block;
497 dprintk("lockd: nlmsvc_cancel(%s/%ld, pi=%d, %Ld-%Ld)\n",
498 file->f_file->f_dentry->d_inode->i_sb->s_id,
499 file->f_file->f_dentry->d_inode->i_ino,
501 (long long)lock->fl.fl_start,
502 (long long)lock->fl.fl_end);
505 block = nlmsvc_lookup_block(file, lock);
508 status = nlmsvc_unlink_block(block);
509 nlmsvc_release_block(block);
511 return status ? nlm_lck_denied : nlm_granted;
515 * Unblock a blocked lock request. This is a callback invoked from the
516 * VFS layer when a lock on which we blocked is removed.
518 * This function doesn't grant the blocked lock instantly, but rather moves
519 * the block to the head of nlm_blocked where it can be picked up by lockd.
522 nlmsvc_notify_blocked(struct file_lock *fl)
524 struct nlm_block *block;
526 dprintk("lockd: VFS unblock notification for block %p\n", fl);
527 list_for_each_entry(block, &nlm_blocked, b_list) {
528 if (nlm_compare_locks(&block->b_call->a_args.lock.fl, fl)) {
529 nlmsvc_insert_block(block, 0);
530 svc_wake_up(block->b_daemon);
535 printk(KERN_WARNING "lockd: notification for unknown block!\n");
538 static int nlmsvc_same_owner(struct file_lock *fl1, struct file_lock *fl2)
540 return fl1->fl_owner == fl2->fl_owner && fl1->fl_pid == fl2->fl_pid;
543 struct lock_manager_operations nlmsvc_lock_operations = {
544 .fl_compare_owner = nlmsvc_same_owner,
545 .fl_notify = nlmsvc_notify_blocked,
549 * Try to claim a lock that was previously blocked.
551 * Note that we use both the RPC_GRANTED_MSG call _and_ an async
552 * RPC thread when notifying the client. This seems like overkill...
554 * - we don't want to use a synchronous RPC thread, otherwise
555 * we might find ourselves hanging on a dead portmapper.
556 * - Some lockd implementations (e.g. HP) don't react to
557 * RPC_GRANTED calls; they seem to insist on RPC_GRANTED_MSG calls.
560 nlmsvc_grant_blocked(struct nlm_block *block)
562 struct nlm_file *file = block->b_file;
563 struct nlm_lock *lock = &block->b_call->a_args.lock;
566 dprintk("lockd: grant blocked lock %p\n", block);
568 /* Unlink block request from list */
569 nlmsvc_unlink_block(block);
571 /* If b_granted is true this means we've been here before.
572 * Just retry the grant callback, possibly refreshing the RPC
574 if (block->b_granted) {
575 nlm_rebind_host(block->b_host);
579 /* Try the lock operation again */
580 lock->fl.fl_flags |= FL_SLEEP;
581 error = posix_lock_file(file->f_file, &lock->fl);
582 lock->fl.fl_flags &= ~FL_SLEEP;
588 dprintk("lockd: lock still blocked\n");
589 nlmsvc_insert_block(block, NLM_NEVER);
592 printk(KERN_WARNING "lockd: unexpected error %d in %s!\n",
593 -error, __FUNCTION__);
594 nlmsvc_insert_block(block, 10 * HZ);
599 /* Lock was granted by VFS. */
600 dprintk("lockd: GRANTing blocked lock.\n");
601 block->b_granted = 1;
603 /* Schedule next grant callback in 30 seconds */
604 nlmsvc_insert_block(block, 30 * HZ);
606 /* Call the client */
607 kref_get(&block->b_count);
608 if (nlm_async_call(block->b_call, NLMPROC_GRANTED_MSG,
609 &nlmsvc_grant_ops) < 0)
610 nlmsvc_release_block(block);
614 * This is the callback from the RPC layer when the NLM_GRANTED_MSG
615 * RPC call has succeeded or timed out.
616 * Like all RPC callbacks, it is invoked by the rpciod process, so it
617 * better not sleep. Therefore, we put the blocked lock on the nlm_blocked
618 * chain once more in order to have it removed by lockd itself (which can
619 * then sleep on the file semaphore without disrupting e.g. the nfs client).
621 static void nlmsvc_grant_callback(struct rpc_task *task, void *data)
623 struct nlm_rqst *call = data;
624 struct nlm_block *block = call->a_block;
625 unsigned long timeout;
627 dprintk("lockd: GRANT_MSG RPC callback\n");
629 /* Technically, we should down the file semaphore here. Since we
630 * move the block towards the head of the queue only, no harm
631 * can be done, though. */
632 if (task->tk_status < 0) {
633 /* RPC error: Re-insert for retransmission */
636 /* Call was successful, now wait for client callback */
639 nlmsvc_insert_block(block, timeout);
640 svc_wake_up(block->b_daemon);
643 static void nlmsvc_grant_release(void *data)
645 struct nlm_rqst *call = data;
647 nlmsvc_release_block(call->a_block);
650 static const struct rpc_call_ops nlmsvc_grant_ops = {
651 .rpc_call_done = nlmsvc_grant_callback,
652 .rpc_release = nlmsvc_grant_release,
656 * We received a GRANT_RES callback. Try to find the corresponding
660 nlmsvc_grant_reply(struct svc_rqst *rqstp, struct nlm_cookie *cookie, u32 status)
662 struct nlm_block *block;
663 struct nlm_file *file;
665 dprintk("grant_reply: looking for cookie %x, host (%08x), s=%d \n",
666 *(unsigned int *)(cookie->data),
667 ntohl(rqstp->rq_addr.sin_addr.s_addr), status);
668 if (!(block = nlmsvc_find_block(cookie, &rqstp->rq_addr)))
670 file = block->b_file;
673 if (status == NLM_LCK_DENIED_GRACE_PERIOD) {
674 /* Try again in a couple of seconds */
675 nlmsvc_insert_block(block, 10 * HZ);
677 /* Lock is now held by client, or has been rejected.
678 * In both cases, the block should be removed. */
679 nlmsvc_unlink_block(block);
682 nlmsvc_release_block(block);
686 * Retry all blocked locks that have been notified. This is where lockd
687 * picks up locks that can be granted, or grant notifications that must
691 nlmsvc_retry_blocked(void)
693 unsigned long timeout = MAX_SCHEDULE_TIMEOUT;
694 struct nlm_block *block;
696 while (!list_empty(&nlm_blocked)) {
697 block = list_entry(nlm_blocked.next, struct nlm_block, b_list);
699 if (block->b_when == NLM_NEVER)
701 if (time_after(block->b_when,jiffies)) {
702 timeout = block->b_when - jiffies;
706 dprintk("nlmsvc_retry_blocked(%p, when=%ld)\n",
707 block, block->b_when);
708 kref_get(&block->b_count);
709 nlmsvc_grant_blocked(block);
710 nlmsvc_release_block(block);