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 and optionally remove it from
109 static struct nlm_block *
110 nlmsvc_lookup_block(struct nlm_file *file, struct nlm_lock *lock, int remove)
112 struct nlm_block **head, *block;
113 struct file_lock *fl;
115 dprintk("lockd: nlmsvc_lookup_block f=%p pd=%d %Ld-%Ld ty=%d\n",
116 file, lock->fl.fl_pid,
117 (long long)lock->fl.fl_start,
118 (long long)lock->fl.fl_end, lock->fl.fl_type);
119 for (head = &nlm_blocked; (block = *head) != 0; head = &block->b_next) {
120 fl = &block->b_call.a_args.lock.fl;
121 dprintk("lockd: check f=%p pd=%d %Ld-%Ld ty=%d cookie=%s\n",
122 block->b_file, fl->fl_pid,
123 (long long)fl->fl_start,
124 (long long)fl->fl_end, fl->fl_type,
125 nlmdbg_cookie2a(&block->b_call.a_args.cookie));
126 if (block->b_file == file && nlm_compare_locks(fl, &lock->fl)) {
128 *head = block->b_next;
131 kref_get(&block->b_count);
139 static inline int nlm_cookie_match(struct nlm_cookie *a, struct nlm_cookie *b)
143 if(memcmp(a->data,b->data,a->len))
149 * Find a block with a given NLM cookie.
151 static inline struct nlm_block *
152 nlmsvc_find_block(struct nlm_cookie *cookie, struct sockaddr_in *sin)
154 struct nlm_block *block;
156 for (block = nlm_blocked; block; block = block->b_next) {
157 dprintk("cookie: head of blocked queue %p, block %p\n",
159 if (nlm_cookie_match(&block->b_call.a_args.cookie,cookie)
160 && nlm_cmp_addr(sin, &block->b_host->h_addr))
165 kref_get(&block->b_count);
170 * Create a block and initialize it.
172 * Note: we explicitly set the cookie of the grant reply to that of
173 * the blocked lock request. The spec explicitly mentions that the client
174 * should _not_ rely on the callback containing the same cookie as the
175 * request, but (as I found out later) that's because some implementations
176 * do just this. Never mind the standards comittees, they support our
177 * logging industries.
179 static inline struct nlm_block *
180 nlmsvc_create_block(struct svc_rqst *rqstp, struct nlm_file *file,
181 struct nlm_lock *lock, struct nlm_cookie *cookie)
183 struct nlm_block *block;
184 struct nlm_host *host;
185 struct nlm_rqst *call;
187 /* Create host handle for callback */
188 host = nlmsvc_lookup_host(rqstp);
192 /* Allocate memory for block, and initialize arguments */
193 if (!(block = (struct nlm_block *) kmalloc(sizeof(*block), GFP_KERNEL)))
195 memset(block, 0, sizeof(*block));
196 locks_init_lock(&block->b_call.a_args.lock.fl);
197 locks_init_lock(&block->b_call.a_res.lock.fl);
198 kref_init(&block->b_count);
200 if (!nlmsvc_setgrantargs(&block->b_call, lock))
203 /* Set notifier function for VFS, and init args */
204 block->b_call.a_args.lock.fl.fl_flags |= FL_SLEEP;
205 block->b_call.a_args.lock.fl.fl_lmops = &nlmsvc_lock_operations;
206 block->b_call.a_args.cookie = *cookie; /* see above */
208 dprintk("lockd: created block %p...\n", block);
210 /* Create and initialize the block */
211 block->b_daemon = rqstp->rq_server;
212 block->b_host = host;
213 block->b_file = file;
215 /* Add to file's list of blocks */
216 block->b_fnext = file->f_blocks;
217 file->f_blocks = block;
219 /* Set up RPC arguments for callback */
220 call = &block->b_call;
222 call->a_flags = RPC_TASK_ASYNC;
229 nlm_release_host(host);
234 * Delete a block. If the lock was cancelled or the grant callback
235 * failed, unlock is set to 1.
236 * It is the caller's responsibility to check whether the file
237 * can be closed hereafter.
239 static int nlmsvc_unlink_block(struct nlm_block *block)
242 dprintk("lockd: unlinking block %p...\n", block);
244 /* Remove block from list */
245 status = posix_unblock_lock(block->b_file->f_file, &block->b_call.a_args.lock.fl);
246 nlmsvc_remove_block(block);
250 static void nlmsvc_free_block(struct kref *kref)
252 struct nlm_block *block = container_of(kref, struct nlm_block, b_count);
253 struct nlm_file *file = block->b_file;
254 struct nlm_block **bp;
256 dprintk("lockd: freeing block %p...\n", block);
258 /* Remove block from file's list of blocks */
259 for (bp = &file->f_blocks; *bp; bp = &(*bp)->b_fnext) {
261 *bp = block->b_fnext;
267 nlm_release_host(block->b_host);
268 nlmsvc_freegrantargs(&block->b_call);
272 static void nlmsvc_release_block(struct nlm_block *block)
275 kref_put(&block->b_count, nlmsvc_free_block);
279 * Loop over all blocks and perform the action specified.
280 * (NLM_ACT_CHECK handled by nlmsvc_inspect_file).
283 nlmsvc_traverse_blocks(struct nlm_host *host, struct nlm_file *file, int action)
285 struct nlm_block *block, *next;
286 /* XXX: Will everything get cleaned up if we don't unlock here? */
289 for (block = file->f_blocks; block; block = next) {
290 next = block->b_fnext;
291 if (action == NLM_ACT_MARK)
292 block->b_host->h_inuse = 1;
293 else if (action == NLM_ACT_UNLOCK) {
294 if (host == NULL || host == block->b_host)
295 nlmsvc_unlink_block(block);
303 * Initialize arguments for GRANTED call. The nlm_rqst structure
304 * has been cleared already.
306 static int nlmsvc_setgrantargs(struct nlm_rqst *call, struct nlm_lock *lock)
308 locks_copy_lock(&call->a_args.lock.fl, &lock->fl);
309 memcpy(&call->a_args.lock.fh, &lock->fh, sizeof(call->a_args.lock.fh));
310 call->a_args.lock.caller = system_utsname.nodename;
311 call->a_args.lock.oh.len = lock->oh.len;
313 /* set default data area */
314 call->a_args.lock.oh.data = call->a_owner;
315 call->a_args.lock.svid = lock->fl.fl_pid;
317 if (lock->oh.len > NLMCLNT_OHSIZE) {
318 void *data = kmalloc(lock->oh.len, GFP_KERNEL);
320 nlmsvc_freegrantargs(call);
323 call->a_args.lock.oh.data = (u8 *) data;
326 memcpy(call->a_args.lock.oh.data, lock->oh.data, lock->oh.len);
330 static void nlmsvc_freegrantargs(struct nlm_rqst *call)
332 struct file_lock *fl = &call->a_args.lock.fl;
334 * Check whether we allocated memory for the owner.
336 if (call->a_args.lock.oh.data != (u8 *) call->a_owner) {
337 kfree(call->a_args.lock.oh.data);
339 if (fl->fl_ops && fl->fl_ops->fl_release_private)
340 fl->fl_ops->fl_release_private(fl);
341 if (fl->fl_lmops && fl->fl_lmops->fl_release_private)
342 fl->fl_lmops->fl_release_private(fl);
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, 0);
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->oh.len = 0; /* don't return OH info */
448 conflock->svid = conflock->fl.fl_pid;
449 return nlm_lck_denied;
457 * This implies a CANCEL call: We send a GRANT_MSG, the client replies
458 * with a GRANT_RES call which gets lost, and calls UNLOCK immediately
459 * afterwards. In this case the block will still be there, and hence
463 nlmsvc_unlock(struct nlm_file *file, struct nlm_lock *lock)
467 dprintk("lockd: nlmsvc_unlock(%s/%ld, pi=%d, %Ld-%Ld)\n",
468 file->f_file->f_dentry->d_inode->i_sb->s_id,
469 file->f_file->f_dentry->d_inode->i_ino,
471 (long long)lock->fl.fl_start,
472 (long long)lock->fl.fl_end);
474 /* First, cancel any lock that might be there */
475 nlmsvc_cancel_blocked(file, lock);
477 lock->fl.fl_type = F_UNLCK;
478 error = posix_lock_file(file->f_file, &lock->fl);
480 return (error < 0)? nlm_lck_denied_nolocks : nlm_granted;
484 * Cancel a previously blocked request.
486 * A cancel request always overrides any grant that may currently
488 * The calling procedure must check whether the file can be closed.
491 nlmsvc_cancel_blocked(struct nlm_file *file, struct nlm_lock *lock)
493 struct nlm_block *block;
496 dprintk("lockd: nlmsvc_cancel(%s/%ld, pi=%d, %Ld-%Ld)\n",
497 file->f_file->f_dentry->d_inode->i_sb->s_id,
498 file->f_file->f_dentry->d_inode->i_ino,
500 (long long)lock->fl.fl_start,
501 (long long)lock->fl.fl_end);
504 if ((block = nlmsvc_lookup_block(file, lock, 1)) != NULL) {
505 status = nlmsvc_unlink_block(block);
506 nlmsvc_release_block(block);
509 return status ? nlm_lck_denied : nlm_granted;
513 * Unblock a blocked lock request. This is a callback invoked from the
514 * VFS layer when a lock on which we blocked is removed.
516 * This function doesn't grant the blocked lock instantly, but rather moves
517 * the block to the head of nlm_blocked where it can be picked up by lockd.
520 nlmsvc_notify_blocked(struct file_lock *fl)
522 struct nlm_block **bp, *block;
524 dprintk("lockd: VFS unblock notification for block %p\n", fl);
525 for (bp = &nlm_blocked; (block = *bp) != 0; bp = &block->b_next) {
526 if (nlm_compare_locks(&block->b_call.a_args.lock.fl, fl)) {
527 nlmsvc_insert_block(block, 0);
528 svc_wake_up(block->b_daemon);
533 printk(KERN_WARNING "lockd: notification for unknown block!\n");
536 static int nlmsvc_same_owner(struct file_lock *fl1, struct file_lock *fl2)
538 return fl1->fl_owner == fl2->fl_owner && fl1->fl_pid == fl2->fl_pid;
541 struct lock_manager_operations nlmsvc_lock_operations = {
542 .fl_compare_owner = nlmsvc_same_owner,
543 .fl_notify = nlmsvc_notify_blocked,
547 * Try to claim a lock that was previously blocked.
549 * Note that we use both the RPC_GRANTED_MSG call _and_ an async
550 * RPC thread when notifying the client. This seems like overkill...
552 * - we don't want to use a synchronous RPC thread, otherwise
553 * we might find ourselves hanging on a dead portmapper.
554 * - Some lockd implementations (e.g. HP) don't react to
555 * RPC_GRANTED calls; they seem to insist on RPC_GRANTED_MSG calls.
558 nlmsvc_grant_blocked(struct nlm_block *block)
560 struct nlm_file *file = block->b_file;
561 struct nlm_lock *lock = &block->b_call.a_args.lock;
564 dprintk("lockd: grant blocked lock %p\n", block);
566 /* First thing is lock the file */
569 /* Unlink block request from list */
570 nlmsvc_unlink_block(block);
572 /* If b_granted is true this means we've been here before.
573 * Just retry the grant callback, possibly refreshing the RPC
575 if (block->b_granted) {
576 nlm_rebind_host(block->b_host);
580 /* Try the lock operation again */
581 lock->fl.fl_flags |= FL_SLEEP;
582 error = posix_lock_file(file->f_file, &lock->fl);
583 lock->fl.fl_flags &= ~FL_SLEEP;
589 dprintk("lockd: lock still blocked\n");
590 nlmsvc_insert_block(block, NLM_NEVER);
593 printk(KERN_WARNING "lockd: unexpected error %d in %s!\n",
594 -error, __FUNCTION__);
595 nlmsvc_insert_block(block, 10 * HZ);
600 /* Lock was granted by VFS. */
601 dprintk("lockd: GRANTing blocked lock.\n");
602 block->b_granted = 1;
604 /* Schedule next grant callback in 30 seconds */
605 nlmsvc_insert_block(block, 30 * HZ);
607 /* Call the client */
608 kref_get(&block->b_count);
609 if (nlmsvc_async_call(&block->b_call, NLMPROC_GRANTED_MSG,
610 &nlmsvc_grant_ops) < 0)
611 nlmsvc_release_block(block);
617 * This is the callback from the RPC layer when the NLM_GRANTED_MSG
618 * RPC call has succeeded or timed out.
619 * Like all RPC callbacks, it is invoked by the rpciod process, so it
620 * better not sleep. Therefore, we put the blocked lock on the nlm_blocked
621 * chain once more in order to have it removed by lockd itself (which can
622 * then sleep on the file semaphore without disrupting e.g. the nfs client).
624 static void nlmsvc_grant_callback(struct rpc_task *task, void *data)
626 struct nlm_rqst *call = data;
627 struct nlm_block *block = container_of(call, struct nlm_block, b_call);
628 unsigned long timeout;
630 dprintk("lockd: GRANT_MSG RPC callback\n");
632 /* Technically, we should down the file semaphore here. Since we
633 * move the block towards the head of the queue only, no harm
634 * can be done, though. */
635 if (task->tk_status < 0) {
636 /* RPC error: Re-insert for retransmission */
638 } else if (block->b_done) {
639 /* Block already removed, kill it for real */
642 /* Call was successful, now wait for client callback */
645 nlmsvc_insert_block(block, timeout);
646 svc_wake_up(block->b_daemon);
649 void nlmsvc_grant_release(void *data)
651 nlmsvc_release_block(data);
654 static const struct rpc_call_ops nlmsvc_grant_ops = {
655 .rpc_call_done = nlmsvc_grant_callback,
656 .rpc_release = nlmsvc_grant_release,
660 * We received a GRANT_RES callback. Try to find the corresponding
664 nlmsvc_grant_reply(struct svc_rqst *rqstp, struct nlm_cookie *cookie, u32 status)
666 struct nlm_block *block;
667 struct nlm_file *file;
669 dprintk("grant_reply: looking for cookie %x, host (%08x), s=%d \n",
670 *(unsigned int *)(cookie->data),
671 ntohl(rqstp->rq_addr.sin_addr.s_addr), status);
672 if (!(block = nlmsvc_find_block(cookie, &rqstp->rq_addr)))
674 file = block->b_file;
679 if (status == NLM_LCK_DENIED_GRACE_PERIOD) {
680 /* Try again in a couple of seconds */
681 nlmsvc_insert_block(block, 10 * HZ);
683 /* Lock is now held by client, or has been rejected.
684 * In both cases, the block should be removed. */
685 nlmsvc_unlink_block(block);
689 nlm_release_file(file);
690 nlmsvc_release_block(block);
694 * Retry all blocked locks that have been notified. This is where lockd
695 * picks up locks that can be granted, or grant notifications that must
699 nlmsvc_retry_blocked(void)
701 struct nlm_block *block;
703 dprintk("nlmsvc_retry_blocked(%p, when=%ld)\n",
705 nlm_blocked? nlm_blocked->b_when : 0);
706 while ((block = nlm_blocked) != 0) {
707 if (block->b_when == NLM_NEVER)
709 if (time_after(block->b_when,jiffies))
711 dprintk("nlmsvc_retry_blocked(%p, when=%ld, done=%d)\n",
712 block, block->b_when, block->b_done);
713 kref_get(&block->b_count);
715 nlmsvc_unlink_block(block);
717 nlmsvc_grant_blocked(block);
718 nlmsvc_release_block(block);
721 if ((block = nlm_blocked) && block->b_when != NLM_NEVER)
722 return (block->b_when - jiffies);
724 return MAX_SCHEDULE_TIMEOUT;