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 = NULL;
187 /* Create host handle for callback */
188 host = nlmsvc_lookup_host(rqstp);
192 call = nlm_alloc_call(host);
196 /* Allocate memory for block, and initialize arguments */
197 block = kzalloc(sizeof(*block), GFP_KERNEL);
200 kref_init(&block->b_count);
202 if (!nlmsvc_setgrantargs(call, lock))
205 /* Set notifier function for VFS, and init args */
206 call->a_args.lock.fl.fl_flags |= FL_SLEEP;
207 call->a_args.lock.fl.fl_lmops = &nlmsvc_lock_operations;
208 call->a_args.cookie = *cookie; /* see above */
210 dprintk("lockd: created block %p...\n", block);
212 /* Create and initialize the block */
213 block->b_daemon = rqstp->rq_server;
214 block->b_host = host;
215 block->b_file = file;
217 /* Add to file's list of blocks */
218 block->b_fnext = file->f_blocks;
219 file->f_blocks = block;
221 /* Set up RPC arguments for callback */
222 block->b_call = call;
223 call->a_flags = RPC_TASK_ASYNC;
224 call->a_block = block;
231 nlm_release_call(call);
236 * Delete a block. If the lock was cancelled or the grant callback
237 * failed, unlock is set to 1.
238 * It is the caller's responsibility to check whether the file
239 * can be closed hereafter.
241 static int nlmsvc_unlink_block(struct nlm_block *block)
244 dprintk("lockd: unlinking block %p...\n", block);
246 /* Remove block from list */
247 status = posix_unblock_lock(block->b_file->f_file, &block->b_call->a_args.lock.fl);
248 nlmsvc_remove_block(block);
252 static void nlmsvc_free_block(struct kref *kref)
254 struct nlm_block *block = container_of(kref, struct nlm_block, b_count);
255 struct nlm_file *file = block->b_file;
256 struct nlm_block **bp;
258 dprintk("lockd: freeing block %p...\n", block);
260 /* Remove block from file's list of blocks */
261 for (bp = &file->f_blocks; *bp; bp = &(*bp)->b_fnext) {
263 *bp = block->b_fnext;
268 nlmsvc_freegrantargs(block->b_call);
269 nlm_release_call(block->b_call);
273 static void nlmsvc_release_block(struct nlm_block *block)
276 kref_put(&block->b_count, nlmsvc_free_block);
280 * Loop over all blocks and perform the action specified.
281 * (NLM_ACT_CHECK handled by nlmsvc_inspect_file).
284 nlmsvc_traverse_blocks(struct nlm_host *host, struct nlm_file *file, int action)
286 struct nlm_block *block, *next;
287 /* XXX: Will everything get cleaned up if we don't unlock here? */
290 for (block = file->f_blocks; block; block = next) {
291 next = block->b_fnext;
292 if (action == NLM_ACT_MARK)
293 block->b_host->h_inuse = 1;
294 else if (action == NLM_ACT_UNLOCK) {
295 if (host == NULL || host == block->b_host)
296 nlmsvc_unlink_block(block);
304 * Initialize arguments for GRANTED call. The nlm_rqst structure
305 * has been cleared already.
307 static int nlmsvc_setgrantargs(struct nlm_rqst *call, struct nlm_lock *lock)
309 locks_copy_lock(&call->a_args.lock.fl, &lock->fl);
310 memcpy(&call->a_args.lock.fh, &lock->fh, sizeof(call->a_args.lock.fh));
311 call->a_args.lock.caller = system_utsname.nodename;
312 call->a_args.lock.oh.len = lock->oh.len;
314 /* set default data area */
315 call->a_args.lock.oh.data = call->a_owner;
316 call->a_args.lock.svid = lock->fl.fl_pid;
318 if (lock->oh.len > NLMCLNT_OHSIZE) {
319 void *data = kmalloc(lock->oh.len, GFP_KERNEL);
322 call->a_args.lock.oh.data = (u8 *) data;
325 memcpy(call->a_args.lock.oh.data, lock->oh.data, lock->oh.len);
329 static void nlmsvc_freegrantargs(struct nlm_rqst *call)
331 if (call->a_args.lock.oh.data != call->a_owner)
332 kfree(call->a_args.lock.oh.data);
336 * Attempt to establish a lock, and if it can't be granted, block it
340 nlmsvc_lock(struct svc_rqst *rqstp, struct nlm_file *file,
341 struct nlm_lock *lock, int wait, struct nlm_cookie *cookie)
343 struct nlm_block *block, *newblock = NULL;
347 dprintk("lockd: nlmsvc_lock(%s/%ld, ty=%d, pi=%d, %Ld-%Ld, bl=%d)\n",
348 file->f_file->f_dentry->d_inode->i_sb->s_id,
349 file->f_file->f_dentry->d_inode->i_ino,
350 lock->fl.fl_type, lock->fl.fl_pid,
351 (long long)lock->fl.fl_start,
352 (long long)lock->fl.fl_end,
356 lock->fl.fl_flags &= ~FL_SLEEP;
358 /* Lock file against concurrent access */
360 /* Get existing block (in case client is busy-waiting) */
361 block = nlmsvc_lookup_block(file, lock, 0);
363 if (newblock != NULL)
364 lock = &newblock->b_call->a_args.lock;
366 lock = &block->b_call->a_args.lock;
368 error = posix_lock_file(file->f_file, &lock->fl);
369 lock->fl.fl_flags &= ~FL_SLEEP;
371 dprintk("lockd: posix_lock_file returned %d\n", error);
382 default: /* includes ENOLCK */
383 ret = nlm_lck_denied_nolocks;
387 ret = nlm_lck_denied;
391 ret = nlm_lck_blocked;
395 /* If we don't have a block, create and initialize it. Then
396 * retry because we may have slept in kmalloc. */
397 /* We have to release f_sema as nlmsvc_create_block may try to
398 * to claim it while doing host garbage collection */
399 if (newblock == NULL) {
401 dprintk("lockd: blocking on this lock (allocating).\n");
402 if (!(newblock = nlmsvc_create_block(rqstp, file, lock, cookie)))
403 return nlm_lck_denied_nolocks;
407 /* Append to list of blocked */
408 nlmsvc_insert_block(newblock, NLM_NEVER);
411 nlmsvc_release_block(newblock);
412 nlmsvc_release_block(block);
413 dprintk("lockd: nlmsvc_lock returned %u\n", ret);
418 * Test for presence of a conflicting lock.
421 nlmsvc_testlock(struct nlm_file *file, struct nlm_lock *lock,
422 struct nlm_lock *conflock)
424 dprintk("lockd: nlmsvc_testlock(%s/%ld, ty=%d, %Ld-%Ld)\n",
425 file->f_file->f_dentry->d_inode->i_sb->s_id,
426 file->f_file->f_dentry->d_inode->i_ino,
428 (long long)lock->fl.fl_start,
429 (long long)lock->fl.fl_end);
431 if (posix_test_lock(file->f_file, &lock->fl, &conflock->fl)) {
432 dprintk("lockd: conflicting lock(ty=%d, %Ld-%Ld)\n",
433 conflock->fl.fl_type,
434 (long long)conflock->fl.fl_start,
435 (long long)conflock->fl.fl_end);
436 conflock->caller = "somehost"; /* FIXME */
437 conflock->oh.len = 0; /* don't return OH info */
438 conflock->svid = conflock->fl.fl_pid;
439 return nlm_lck_denied;
447 * This implies a CANCEL call: We send a GRANT_MSG, the client replies
448 * with a GRANT_RES call which gets lost, and calls UNLOCK immediately
449 * afterwards. In this case the block will still be there, and hence
453 nlmsvc_unlock(struct nlm_file *file, struct nlm_lock *lock)
457 dprintk("lockd: nlmsvc_unlock(%s/%ld, pi=%d, %Ld-%Ld)\n",
458 file->f_file->f_dentry->d_inode->i_sb->s_id,
459 file->f_file->f_dentry->d_inode->i_ino,
461 (long long)lock->fl.fl_start,
462 (long long)lock->fl.fl_end);
464 /* First, cancel any lock that might be there */
465 nlmsvc_cancel_blocked(file, lock);
467 lock->fl.fl_type = F_UNLCK;
468 error = posix_lock_file(file->f_file, &lock->fl);
470 return (error < 0)? nlm_lck_denied_nolocks : nlm_granted;
474 * Cancel a previously blocked request.
476 * A cancel request always overrides any grant that may currently
478 * The calling procedure must check whether the file can be closed.
481 nlmsvc_cancel_blocked(struct nlm_file *file, struct nlm_lock *lock)
483 struct nlm_block *block;
486 dprintk("lockd: nlmsvc_cancel(%s/%ld, pi=%d, %Ld-%Ld)\n",
487 file->f_file->f_dentry->d_inode->i_sb->s_id,
488 file->f_file->f_dentry->d_inode->i_ino,
490 (long long)lock->fl.fl_start,
491 (long long)lock->fl.fl_end);
494 if ((block = nlmsvc_lookup_block(file, lock, 1)) != NULL) {
495 status = nlmsvc_unlink_block(block);
496 nlmsvc_release_block(block);
499 return status ? nlm_lck_denied : nlm_granted;
503 * Unblock a blocked lock request. This is a callback invoked from the
504 * VFS layer when a lock on which we blocked is removed.
506 * This function doesn't grant the blocked lock instantly, but rather moves
507 * the block to the head of nlm_blocked where it can be picked up by lockd.
510 nlmsvc_notify_blocked(struct file_lock *fl)
512 struct nlm_block **bp, *block;
514 dprintk("lockd: VFS unblock notification for block %p\n", fl);
515 for (bp = &nlm_blocked; (block = *bp) != 0; bp = &block->b_next) {
516 if (nlm_compare_locks(&block->b_call->a_args.lock.fl, fl)) {
517 nlmsvc_insert_block(block, 0);
518 svc_wake_up(block->b_daemon);
523 printk(KERN_WARNING "lockd: notification for unknown block!\n");
526 static int nlmsvc_same_owner(struct file_lock *fl1, struct file_lock *fl2)
528 return fl1->fl_owner == fl2->fl_owner && fl1->fl_pid == fl2->fl_pid;
531 struct lock_manager_operations nlmsvc_lock_operations = {
532 .fl_compare_owner = nlmsvc_same_owner,
533 .fl_notify = nlmsvc_notify_blocked,
537 * Try to claim a lock that was previously blocked.
539 * Note that we use both the RPC_GRANTED_MSG call _and_ an async
540 * RPC thread when notifying the client. This seems like overkill...
542 * - we don't want to use a synchronous RPC thread, otherwise
543 * we might find ourselves hanging on a dead portmapper.
544 * - Some lockd implementations (e.g. HP) don't react to
545 * RPC_GRANTED calls; they seem to insist on RPC_GRANTED_MSG calls.
548 nlmsvc_grant_blocked(struct nlm_block *block)
550 struct nlm_file *file = block->b_file;
551 struct nlm_lock *lock = &block->b_call->a_args.lock;
554 dprintk("lockd: grant blocked lock %p\n", block);
556 /* First thing is lock the file */
559 /* Unlink block request from list */
560 nlmsvc_unlink_block(block);
562 /* If b_granted is true this means we've been here before.
563 * Just retry the grant callback, possibly refreshing the RPC
565 if (block->b_granted) {
566 nlm_rebind_host(block->b_host);
570 /* Try the lock operation again */
571 lock->fl.fl_flags |= FL_SLEEP;
572 error = posix_lock_file(file->f_file, &lock->fl);
573 lock->fl.fl_flags &= ~FL_SLEEP;
579 dprintk("lockd: lock still blocked\n");
580 nlmsvc_insert_block(block, NLM_NEVER);
583 printk(KERN_WARNING "lockd: unexpected error %d in %s!\n",
584 -error, __FUNCTION__);
585 nlmsvc_insert_block(block, 10 * HZ);
590 /* Lock was granted by VFS. */
591 dprintk("lockd: GRANTing blocked lock.\n");
592 block->b_granted = 1;
594 /* Schedule next grant callback in 30 seconds */
595 nlmsvc_insert_block(block, 30 * HZ);
597 /* Call the client */
598 kref_get(&block->b_count);
599 if (nlm_async_call(block->b_call, NLMPROC_GRANTED_MSG,
600 &nlmsvc_grant_ops) < 0)
601 nlmsvc_release_block(block);
607 * This is the callback from the RPC layer when the NLM_GRANTED_MSG
608 * RPC call has succeeded or timed out.
609 * Like all RPC callbacks, it is invoked by the rpciod process, so it
610 * better not sleep. Therefore, we put the blocked lock on the nlm_blocked
611 * chain once more in order to have it removed by lockd itself (which can
612 * then sleep on the file semaphore without disrupting e.g. the nfs client).
614 static void nlmsvc_grant_callback(struct rpc_task *task, void *data)
616 struct nlm_rqst *call = data;
617 struct nlm_block *block = call->a_block;
618 unsigned long timeout;
620 dprintk("lockd: GRANT_MSG RPC callback\n");
622 /* Technically, we should down the file semaphore here. Since we
623 * move the block towards the head of the queue only, no harm
624 * can be done, though. */
625 if (task->tk_status < 0) {
626 /* RPC error: Re-insert for retransmission */
628 } else if (block->b_done) {
629 /* Block already removed, kill it for real */
632 /* Call was successful, now wait for client callback */
635 nlmsvc_insert_block(block, timeout);
636 svc_wake_up(block->b_daemon);
639 void nlmsvc_grant_release(void *data)
641 struct nlm_rqst *call = data;
643 nlmsvc_release_block(call->a_block);
646 static const struct rpc_call_ops nlmsvc_grant_ops = {
647 .rpc_call_done = nlmsvc_grant_callback,
648 .rpc_release = nlmsvc_grant_release,
652 * We received a GRANT_RES callback. Try to find the corresponding
656 nlmsvc_grant_reply(struct svc_rqst *rqstp, struct nlm_cookie *cookie, u32 status)
658 struct nlm_block *block;
659 struct nlm_file *file;
661 dprintk("grant_reply: looking for cookie %x, host (%08x), s=%d \n",
662 *(unsigned int *)(cookie->data),
663 ntohl(rqstp->rq_addr.sin_addr.s_addr), status);
664 if (!(block = nlmsvc_find_block(cookie, &rqstp->rq_addr)))
666 file = block->b_file;
671 if (status == NLM_LCK_DENIED_GRACE_PERIOD) {
672 /* Try again in a couple of seconds */
673 nlmsvc_insert_block(block, 10 * HZ);
675 /* Lock is now held by client, or has been rejected.
676 * In both cases, the block should be removed. */
677 nlmsvc_unlink_block(block);
681 nlm_release_file(file);
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 struct nlm_block *block;
695 dprintk("nlmsvc_retry_blocked(%p, when=%ld)\n",
697 nlm_blocked? nlm_blocked->b_when : 0);
698 while ((block = nlm_blocked) != 0) {
699 if (block->b_when == NLM_NEVER)
701 if (time_after(block->b_when,jiffies))
703 dprintk("nlmsvc_retry_blocked(%p, when=%ld, done=%d)\n",
704 block, block->b_when, block->b_done);
705 kref_get(&block->b_count);
707 nlmsvc_unlink_block(block);
709 nlmsvc_grant_blocked(block);
710 nlmsvc_release_block(block);
713 if ((block = nlm_blocked) && block->b_when != NLM_NEVER)
714 return (block->b_when - jiffies);
716 return MAX_SCHEDULE_TIMEOUT;