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_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);
45 static void nlmsvc_notify_blocked(struct file_lock *);
48 * The list of blocked locks to retry
50 static struct nlm_block * nlm_blocked;
53 * Insert a blocked lock into the global list
56 nlmsvc_insert_block(struct nlm_block *block, unsigned long when)
58 struct nlm_block **bp, *b;
60 dprintk("lockd: nlmsvc_insert_block(%p, %ld)\n", block, when);
62 nlmsvc_remove_block(block);
64 if (when != NLM_NEVER) {
65 if ((when += jiffies) == NLM_NEVER)
67 while ((b = *bp) && time_before_eq(b->b_when,when) && b->b_when != NLM_NEVER)
80 * Remove a block from the global list
83 nlmsvc_remove_block(struct nlm_block *block)
85 struct nlm_block **bp, *b;
89 for (bp = &nlm_blocked; (b = *bp); bp = &b->b_next) {
101 * Find a block for a given lock and optionally remove it from
104 static struct nlm_block *
105 nlmsvc_lookup_block(struct nlm_file *file, struct nlm_lock *lock, int remove)
107 struct nlm_block **head, *block;
108 struct file_lock *fl;
110 dprintk("lockd: nlmsvc_lookup_block f=%p pd=%d %Ld-%Ld ty=%d\n",
111 file, lock->fl.fl_pid,
112 (long long)lock->fl.fl_start,
113 (long long)lock->fl.fl_end, lock->fl.fl_type);
114 for (head = &nlm_blocked; (block = *head); head = &block->b_next) {
115 fl = &block->b_call.a_args.lock.fl;
116 dprintk("lockd: check f=%p pd=%d %Ld-%Ld ty=%d cookie=%s\n",
117 block->b_file, fl->fl_pid,
118 (long long)fl->fl_start,
119 (long long)fl->fl_end, fl->fl_type,
120 nlmdbg_cookie2a(&block->b_call.a_args.cookie));
121 if (block->b_file == file && nlm_compare_locks(fl, &lock->fl)) {
123 *head = block->b_next;
133 static inline int nlm_cookie_match(struct nlm_cookie *a, struct nlm_cookie *b)
137 if(memcmp(a->data,b->data,a->len))
143 * Find a block with a given NLM cookie.
145 static inline struct nlm_block *
146 nlmsvc_find_block(struct nlm_cookie *cookie, struct sockaddr_in *sin)
148 struct nlm_block *block;
150 for (block = nlm_blocked; block; block = block->b_next) {
151 dprintk("cookie: head of blocked queue %p, block %p\n",
153 if (nlm_cookie_match(&block->b_call.a_args.cookie,cookie)
154 && nlm_cmp_addr(sin, &block->b_host->h_addr))
162 * Create a block and initialize it.
164 * Note: we explicitly set the cookie of the grant reply to that of
165 * the blocked lock request. The spec explicitly mentions that the client
166 * should _not_ rely on the callback containing the same cookie as the
167 * request, but (as I found out later) that's because some implementations
168 * do just this. Never mind the standards comittees, they support our
169 * logging industries.
171 static inline struct nlm_block *
172 nlmsvc_create_block(struct svc_rqst *rqstp, struct nlm_file *file,
173 struct nlm_lock *lock, struct nlm_cookie *cookie)
175 struct nlm_block *block;
176 struct nlm_host *host;
177 struct nlm_rqst *call;
179 /* Create host handle for callback */
180 host = nlmclnt_lookup_host(&rqstp->rq_addr,
181 rqstp->rq_prot, rqstp->rq_vers);
185 /* Allocate memory for block, and initialize arguments */
186 if (!(block = (struct nlm_block *) kmalloc(sizeof(*block), GFP_KERNEL)))
188 memset(block, 0, sizeof(*block));
189 locks_init_lock(&block->b_call.a_args.lock.fl);
190 locks_init_lock(&block->b_call.a_res.lock.fl);
192 block->b_host = nlmsvc_lookup_host(rqstp);
193 if (block->b_host == NULL) {
197 if (!nlmclnt_setgrantargs(&block->b_call, lock))
200 /* Set notifier function for VFS, and init args */
201 block->b_call.a_args.lock.fl.fl_notify = nlmsvc_notify_blocked;
202 block->b_call.a_args.cookie = *cookie; /* see above */
204 dprintk("lockd: created block %p...\n", block);
206 /* Create and initialize the block */
207 block->b_daemon = rqstp->rq_server;
208 block->b_file = file;
210 /* Add to file's list of blocks */
211 block->b_fnext = file->f_blocks;
212 file->f_blocks = block;
214 /* Set up RPC arguments for callback */
215 call = &block->b_call;
217 call->a_flags = RPC_TASK_ASYNC;
224 nlm_release_host(host);
229 * Delete a block. If the lock was cancelled or the grant callback
230 * failed, unlock is set to 1.
231 * It is the caller's responsibility to check whether the file
232 * can be closed hereafter.
235 nlmsvc_delete_block(struct nlm_block *block, int unlock)
237 struct file_lock *fl = &block->b_call.a_args.lock.fl;
238 struct nlm_file *file = block->b_file;
239 struct nlm_block **bp;
241 dprintk("lockd: deleting block %p...\n", block);
243 /* Remove block from list */
244 nlmsvc_remove_block(block);
246 /* If granted, unlock it, else remove from inode block list */
247 if (unlock && block->b_granted) {
248 dprintk("lockd: deleting granted lock\n");
249 fl->fl_type = F_UNLCK;
250 posix_lock_file(&block->b_file->f_file, fl, 0);
251 block->b_granted = 0;
253 dprintk("lockd: unblocking blocked lock\n");
254 posix_unblock_lock(fl);
257 /* If the block is in the middle of a GRANT callback,
258 * don't kill it yet. */
259 if (block->b_incall) {
260 nlmsvc_insert_block(block, NLM_NEVER);
265 /* Remove block from file's list of blocks */
266 for (bp = &file->f_blocks; *bp; bp = &(*bp)->b_fnext) {
268 *bp = block->b_fnext;
273 nlm_release_host(block->b_host);
274 nlmclnt_freegrantargs(&block->b_call);
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;
288 for (block = file->f_blocks; block; block = next) {
289 next = block->b_fnext;
290 if (action == NLM_ACT_MARK)
291 block->b_host->h_inuse = 1;
292 else if (action == NLM_ACT_UNLOCK) {
293 if (host == NULL || host == block->b_host)
294 nlmsvc_delete_block(block, 1);
302 * Attempt to establish a lock, and if it can't be granted, block it
306 nlmsvc_lock(struct svc_rqst *rqstp, struct nlm_file *file,
307 struct nlm_lock *lock, int wait, struct nlm_cookie *cookie)
309 struct file_lock *conflock;
310 struct nlm_block *block;
313 dprintk("lockd: nlmsvc_lock(%04x/%ld, ty=%d, pi=%d, %Ld-%Ld, bl=%d)\n",
314 file->f_file.f_dentry->d_inode->i_dev,
315 file->f_file.f_dentry->d_inode->i_ino,
316 lock->fl.fl_type, lock->fl.fl_pid,
317 (long long)lock->fl.fl_start,
318 (long long)lock->fl.fl_end,
322 /* Get existing block (in case client is busy-waiting) */
323 block = nlmsvc_lookup_block(file, lock, 0);
325 lock->fl.fl_flags |= FL_LOCKD;
328 /* Lock file against concurrent access */
331 if (!(conflock = posix_test_lock(&file->f_file, &lock->fl))) {
332 error = posix_lock_file(&file->f_file, &lock->fl, 0);
335 nlmsvc_delete_block(block, 0);
338 dprintk("lockd: posix_lock_file returned %d\n", -error);
345 return nlm_lck_denied;
346 default: /* includes ENOLCK */
347 return nlm_lck_denied_nolocks;
353 return nlm_lck_denied;
356 if (posix_locks_deadlock(&lock->fl, conflock)) {
361 /* If we don't have a block, create and initialize it. Then
362 * retry because we may have slept in kmalloc. */
363 /* We have to release f_sema as nlmsvc_create_block may try to
364 * claim it while doing host garbage collection */
367 dprintk("lockd: blocking on this lock (allocating).\n");
368 if (!(block = nlmsvc_create_block(rqstp, file, lock, cookie)))
369 return nlm_lck_denied_nolocks;
373 /* Append to list of blocked */
374 nlmsvc_insert_block(block, NLM_NEVER);
376 if (list_empty(&block->b_call.a_args.lock.fl.fl_block)) {
377 /* Now add block to block list of the conflicting lock
378 if we haven't done so. */
379 dprintk("lockd: blocking on this lock.\n");
380 posix_block_lock(conflock, &block->b_call.a_args.lock.fl);
384 return nlm_lck_blocked;
388 * Test for presence of a conflicting lock.
391 nlmsvc_testlock(struct nlm_file *file, struct nlm_lock *lock,
392 struct nlm_lock *conflock)
394 struct file_lock *fl;
396 dprintk("lockd: nlmsvc_testlock(%04x/%ld, ty=%d, %Ld-%Ld)\n",
397 file->f_file.f_dentry->d_inode->i_dev,
398 file->f_file.f_dentry->d_inode->i_ino,
400 (long long)lock->fl.fl_start,
401 (long long)lock->fl.fl_end);
403 if ((fl = posix_test_lock(&file->f_file, &lock->fl)) != NULL) {
404 dprintk("lockd: conflicting lock(ty=%d, %Ld-%Ld)\n",
405 fl->fl_type, (long long)fl->fl_start,
406 (long long)fl->fl_end);
407 conflock->caller = "somehost"; /* FIXME */
408 conflock->oh.len = 0; /* don't return OH info */
410 return nlm_lck_denied;
418 * This implies a CANCEL call: We send a GRANT_MSG, the client replies
419 * with a GRANT_RES call which gets lost, and calls UNLOCK immediately
420 * afterwards. In this case the block will still be there, and hence
424 nlmsvc_unlock(struct nlm_file *file, struct nlm_lock *lock)
428 dprintk("lockd: nlmsvc_unlock(%04x/%ld, pi=%d, %Ld-%Ld)\n",
429 file->f_file.f_dentry->d_inode->i_dev,
430 file->f_file.f_dentry->d_inode->i_ino,
432 (long long)lock->fl.fl_start,
433 (long long)lock->fl.fl_end);
435 /* First, cancel any lock that might be there */
436 nlmsvc_cancel_blocked(file, lock);
438 lock->fl.fl_type = F_UNLCK;
439 error = posix_lock_file(&file->f_file, &lock->fl, 0);
441 return (error < 0)? nlm_lck_denied_nolocks : nlm_granted;
445 * Cancel a previously blocked request.
447 * A cancel request always overrides any grant that may currently
449 * The calling procedure must check whether the file can be closed.
452 nlmsvc_cancel_blocked(struct nlm_file *file, struct nlm_lock *lock)
454 struct nlm_block *block;
456 dprintk("lockd: nlmsvc_cancel(%04x/%ld, pi=%d, %Ld-%Ld)\n",
457 file->f_file.f_dentry->d_inode->i_dev,
458 file->f_file.f_dentry->d_inode->i_ino,
460 (long long)lock->fl.fl_start,
461 (long long)lock->fl.fl_end);
464 if ((block = nlmsvc_lookup_block(file, lock, 1)) != NULL)
465 nlmsvc_delete_block(block, 1);
471 * Unblock a blocked lock request. This is a callback invoked from the
472 * VFS layer when a lock on which we blocked is removed.
474 * This function doesn't grant the blocked lock instantly, but rather moves
475 * the block to the head of nlm_blocked where it can be picked up by lockd.
478 nlmsvc_notify_blocked(struct file_lock *fl)
480 struct nlm_block **bp, *block;
482 dprintk("lockd: VFS unblock notification for block %p\n", fl);
483 posix_unblock_lock(fl);
484 for (bp = &nlm_blocked; (block = *bp); bp = &block->b_next) {
485 if (nlm_compare_locks(&block->b_call.a_args.lock.fl, fl)) {
486 nlmsvc_insert_block(block, 0);
487 svc_wake_up(block->b_daemon);
492 printk(KERN_WARNING "lockd: notification for unknown block!\n");
496 * Try to claim a lock that was previously blocked.
498 * Note that we use both the RPC_GRANTED_MSG call _and_ an async
499 * RPC thread when notifying the client. This seems like overkill...
501 * - we don't want to use a synchronous RPC thread, otherwise
502 * we might find ourselves hanging on a dead portmapper.
503 * - Some lockd implementations (e.g. HP) don't react to
504 * RPC_GRANTED calls; they seem to insist on RPC_GRANTED_MSG calls.
507 nlmsvc_grant_blocked(struct nlm_block *block)
509 struct nlm_file *file = block->b_file;
510 struct nlm_lock *lock = &block->b_call.a_args.lock;
511 struct file_lock *conflock;
514 dprintk("lockd: grant blocked lock %p\n", block);
516 /* First thing is lock the file */
519 /* Unlink block request from list */
520 nlmsvc_remove_block(block);
522 /* If b_granted is true this means we've been here before.
523 * Just retry the grant callback, possibly refreshing the RPC
525 if (block->b_granted) {
526 nlm_rebind_host(block->b_call.a_host);
530 /* Try the lock operation again */
531 if ((conflock = posix_test_lock(&file->f_file, &lock->fl)) != NULL) {
532 /* Bummer, we blocked again */
533 dprintk("lockd: lock still blocked\n");
534 nlmsvc_insert_block(block, NLM_NEVER);
535 posix_block_lock(conflock, &lock->fl);
540 /* Alright, no conflicting lock. Now lock it for real. If the
541 * following yields an error, this is most probably due to low
542 * memory. Retry the lock in a few seconds.
544 if ((error = posix_lock_file(&file->f_file, &lock->fl, 0)) < 0) {
545 printk(KERN_WARNING "lockd: unexpected error %d in %s!\n",
546 -error, __FUNCTION__);
547 nlmsvc_insert_block(block, 10 * HZ);
553 /* Lock was granted by VFS. */
554 dprintk("lockd: GRANTing blocked lock.\n");
555 block->b_granted = 1;
558 /* Schedule next grant callback in 30 seconds */
559 nlmsvc_insert_block(block, 30 * HZ);
561 /* Call the client */
562 nlm_get_host(block->b_call.a_host);
563 if (nlmsvc_async_call(&block->b_call, NLMPROC_GRANTED_MSG,
564 nlmsvc_grant_callback) < 0)
565 nlm_release_host(block->b_call.a_host);
570 * This is the callback from the RPC layer when the NLM_GRANTED_MSG
571 * RPC call has succeeded or timed out.
572 * Like all RPC callbacks, it is invoked by the rpciod process, so it
573 * better not sleep. Therefore, we put the blocked lock on the nlm_blocked
574 * chain once more in order to have it removed by lockd itself (which can
575 * then sleep on the file semaphore without disrupting e.g. the nfs client).
578 nlmsvc_grant_callback(struct rpc_task *task)
580 struct nlm_rqst *call = (struct nlm_rqst *) task->tk_calldata;
581 struct nlm_block *block;
582 unsigned long timeout;
583 struct sockaddr_in *peer_addr = RPC_PEERADDR(task->tk_client);
585 dprintk("lockd: GRANT_MSG RPC callback\n");
586 dprintk("callback: looking for cookie %s, host %u.%u.%u.%u\n",
587 nlmdbg_cookie2a(&call->a_args.cookie),
588 NIPQUAD(peer_addr->sin_addr.s_addr));
589 if (!(block = nlmsvc_find_block(&call->a_args.cookie, peer_addr))) {
590 dprintk("lockd: no block for cookie %s, host %u.%u.%u.%u\n",
591 nlmdbg_cookie2a(&call->a_args.cookie),
592 NIPQUAD(peer_addr->sin_addr.s_addr));
596 /* Technically, we should down the file semaphore here. Since we
597 * move the block towards the head of the queue only, no harm
598 * can be done, though. */
599 if (task->tk_status < 0) {
600 /* RPC error: Re-insert for retransmission */
602 } else if (block->b_done) {
603 /* Block already removed, kill it for real */
606 /* Call was successful, now wait for client callback */
609 nlmsvc_insert_block(block, timeout);
610 svc_wake_up(block->b_daemon);
613 nlm_release_host(call->a_host);
617 * We received a GRANT_RES callback. Try to find the corresponding
621 nlmsvc_grant_reply(struct svc_rqst *rqstp, struct nlm_cookie *cookie, u32 status)
623 struct nlm_block *block;
624 struct nlm_file *file;
626 dprintk("grant_reply: looking for cookie %x, host (%08x), s=%d \n",
627 *(unsigned int *)(cookie->data),
628 ntohl(rqstp->rq_addr.sin_addr.s_addr), status);
629 if (!(block = nlmsvc_find_block(cookie, &rqstp->rq_addr)))
631 file = block->b_file;
635 if ((block = nlmsvc_find_block(cookie,&rqstp->rq_addr)) != NULL) {
636 if (status == NLM_LCK_DENIED_GRACE_PERIOD) {
637 /* Try again in a couple of seconds */
638 nlmsvc_insert_block(block, 10 * HZ);
641 /* Lock is now held by client, or has been rejected.
642 * In both cases, the block should be removed. */
644 if (status == NLM_LCK_GRANTED)
645 nlmsvc_delete_block(block, 0);
647 nlmsvc_delete_block(block, 1);
652 nlm_release_file(file);
656 * Retry all blocked locks that have been notified. This is where lockd
657 * picks up locks that can be granted, or grant notifications that must
661 nlmsvc_retry_blocked(void)
663 struct nlm_block *block;
665 dprintk("nlmsvc_retry_blocked(%p, when=%ld)\n",
667 nlm_blocked? nlm_blocked->b_when : 0);
668 while ((block = nlm_blocked)) {
669 if (block->b_when == NLM_NEVER)
671 if (time_after(block->b_when,jiffies))
673 dprintk("nlmsvc_retry_blocked(%p, when=%ld, done=%d)\n",
674 block, block->b_when, block->b_done);
676 nlmsvc_delete_block(block, 0);
678 nlmsvc_grant_blocked(block);
681 if ((block = nlm_blocked) && block->b_when != NLM_NEVER)
682 return (block->b_when - jiffies);
684 return MAX_SCHEDULE_TIMEOUT;