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=%x\n",
117 block->b_file, fl->fl_pid,
118 (long long)fl->fl_start,
119 (long long)fl->fl_end, fl->fl_type,
120 *(unsigned int*)(block->b_call.a_args.cookie.data));
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 if (!nlmclnt_setgrantargs(&block->b_call, lock))
195 /* Set notifier function for VFS, and init args */
196 block->b_call.a_args.lock.fl.fl_notify = nlmsvc_notify_blocked;
197 block->b_call.a_args.cookie = *cookie; /* see above */
199 dprintk("lockd: created block %p...\n", block);
201 /* Create and initialize the block */
202 block->b_daemon = rqstp->rq_server;
203 block->b_host = host;
204 block->b_file = file;
206 /* Add to file's list of blocks */
207 block->b_fnext = file->f_blocks;
208 file->f_blocks = block;
210 /* Set up RPC arguments for callback */
211 call = &block->b_call;
213 call->a_flags = RPC_TASK_ASYNC;
220 nlm_release_host(host);
225 * Delete a block. If the lock was cancelled or the grant callback
226 * failed, unlock is set to 1.
227 * It is the caller's responsibility to check whether the file
228 * can be closed hereafter.
231 nlmsvc_delete_block(struct nlm_block *block, int unlock)
233 struct file_lock *fl = &block->b_call.a_args.lock.fl;
234 struct nlm_file *file = block->b_file;
235 struct nlm_block **bp;
237 dprintk("lockd: deleting block %p...\n", block);
239 /* Remove block from list */
240 nlmsvc_remove_block(block);
241 posix_unblock_lock(&file->f_file, fl);
242 block->b_granted = 0;
244 /* If the block is in the middle of a GRANT callback,
245 * don't kill it yet. */
246 if (block->b_incall) {
247 nlmsvc_insert_block(block, NLM_NEVER);
252 /* Remove block from file's list of blocks */
253 for (bp = &file->f_blocks; *bp; bp = &(*bp)->b_fnext) {
255 *bp = block->b_fnext;
261 nlm_release_host(block->b_host);
262 nlmclnt_freegrantargs(&block->b_call);
267 * Loop over all blocks and perform the action specified.
268 * (NLM_ACT_CHECK handled by nlmsvc_inspect_file).
271 nlmsvc_traverse_blocks(struct nlm_host *host, struct nlm_file *file, int action)
273 struct nlm_block *block, *next;
276 for (block = file->f_blocks; block; block = next) {
277 next = block->b_fnext;
278 if (action == NLM_ACT_MARK)
279 block->b_host->h_inuse = 1;
280 else if (action == NLM_ACT_UNLOCK) {
281 if (host == NULL || host == block->b_host)
282 nlmsvc_delete_block(block, 1);
290 * Attempt to establish a lock, and if it can't be granted, block it
294 nlmsvc_lock(struct svc_rqst *rqstp, struct nlm_file *file,
295 struct nlm_lock *lock, int wait, struct nlm_cookie *cookie)
297 struct file_lock *conflock;
298 struct nlm_block *block;
301 dprintk("lockd: nlmsvc_lock(%s/%ld, ty=%d, pi=%d, %Ld-%Ld, bl=%d)\n",
302 file->f_file.f_dentry->d_inode->i_sb->s_id,
303 file->f_file.f_dentry->d_inode->i_ino,
304 lock->fl.fl_type, lock->fl.fl_pid,
305 (long long)lock->fl.fl_start,
306 (long long)lock->fl.fl_end,
310 /* Get existing block (in case client is busy-waiting) */
311 block = nlmsvc_lookup_block(file, lock, 0);
313 lock->fl.fl_flags |= FL_LOCKD;
316 /* Lock file against concurrent access */
319 if (!(conflock = posix_test_lock(&file->f_file, &lock->fl))) {
320 error = posix_lock_file(&file->f_file, &lock->fl);
323 nlmsvc_delete_block(block, 0);
326 dprintk("lockd: posix_lock_file returned %d\n", -error);
333 return nlm_lck_denied;
334 default: /* includes ENOLCK */
335 return nlm_lck_denied_nolocks;
341 return nlm_lck_denied;
344 if (posix_locks_deadlock(&lock->fl, conflock)) {
349 /* If we don't have a block, create and initialize it. Then
350 * retry because we may have slept in kmalloc. */
351 /* We have to release f_sema as nlmsvc_create_block may try to
352 * to claim it while doing host garbage collection */
355 dprintk("lockd: blocking on this lock (allocating).\n");
356 if (!(block = nlmsvc_create_block(rqstp, file, lock, cookie)))
357 return nlm_lck_denied_nolocks;
361 /* Append to list of blocked */
362 nlmsvc_insert_block(block, NLM_NEVER);
364 if (list_empty(&block->b_call.a_args.lock.fl.fl_block)) {
365 /* Now add block to block list of the conflicting lock
366 if we haven't done so. */
367 dprintk("lockd: blocking on this lock.\n");
368 posix_block_lock(conflock, &block->b_call.a_args.lock.fl);
372 return nlm_lck_blocked;
376 * Test for presence of a conflicting lock.
379 nlmsvc_testlock(struct nlm_file *file, struct nlm_lock *lock,
380 struct nlm_lock *conflock)
382 struct file_lock *fl;
384 dprintk("lockd: nlmsvc_testlock(%s/%ld, ty=%d, %Ld-%Ld)\n",
385 file->f_file.f_dentry->d_inode->i_sb->s_id,
386 file->f_file.f_dentry->d_inode->i_ino,
388 (long long)lock->fl.fl_start,
389 (long long)lock->fl.fl_end);
391 if ((fl = posix_test_lock(&file->f_file, &lock->fl)) != NULL) {
392 dprintk("lockd: conflicting lock(ty=%d, %Ld-%Ld)\n",
393 fl->fl_type, (long long)fl->fl_start,
394 (long long)fl->fl_end);
395 conflock->caller = "somehost"; /* FIXME */
396 conflock->oh.len = 0; /* don't return OH info */
398 return nlm_lck_denied;
406 * This implies a CANCEL call: We send a GRANT_MSG, the client replies
407 * with a GRANT_RES call which gets lost, and calls UNLOCK immediately
408 * afterwards. In this case the block will still be there, and hence
412 nlmsvc_unlock(struct nlm_file *file, struct nlm_lock *lock)
416 dprintk("lockd: nlmsvc_unlock(%s/%ld, pi=%d, %Ld-%Ld)\n",
417 file->f_file.f_dentry->d_inode->i_sb->s_id,
418 file->f_file.f_dentry->d_inode->i_ino,
420 (long long)lock->fl.fl_start,
421 (long long)lock->fl.fl_end);
423 /* First, cancel any lock that might be there */
424 nlmsvc_cancel_blocked(file, lock);
426 lock->fl.fl_type = F_UNLCK;
427 error = posix_lock_file(&file->f_file, &lock->fl);
429 return (error < 0)? nlm_lck_denied_nolocks : nlm_granted;
433 * Cancel a previously blocked request.
435 * A cancel request always overrides any grant that may currently
437 * The calling procedure must check whether the file can be closed.
440 nlmsvc_cancel_blocked(struct nlm_file *file, struct nlm_lock *lock)
442 struct nlm_block *block;
444 dprintk("lockd: nlmsvc_cancel(%s/%ld, pi=%d, %Ld-%Ld)\n",
445 file->f_file.f_dentry->d_inode->i_sb->s_id,
446 file->f_file.f_dentry->d_inode->i_ino,
448 (long long)lock->fl.fl_start,
449 (long long)lock->fl.fl_end);
452 if ((block = nlmsvc_lookup_block(file, lock, 1)) != NULL)
453 nlmsvc_delete_block(block, 1);
459 * Unblock a blocked lock request. This is a callback invoked from the
460 * VFS layer when a lock on which we blocked is removed.
462 * This function doesn't grant the blocked lock instantly, but rather moves
463 * the block to the head of nlm_blocked where it can be picked up by lockd.
466 nlmsvc_notify_blocked(struct file_lock *fl)
468 struct nlm_block **bp, *block;
470 dprintk("lockd: VFS unblock notification for block %p\n", fl);
471 for (bp = &nlm_blocked; (block = *bp); bp = &block->b_next) {
472 if (nlm_compare_locks(&block->b_call.a_args.lock.fl, fl)) {
473 nlmsvc_insert_block(block, 0);
474 svc_wake_up(block->b_daemon);
479 printk(KERN_WARNING "lockd: notification for unknown block!\n");
483 * Try to claim a lock that was previously blocked.
485 * Note that we use both the RPC_GRANTED_MSG call _and_ an async
486 * RPC thread when notifying the client. This seems like overkill...
488 * - we don't want to use a synchronous RPC thread, otherwise
489 * we might find ourselves hanging on a dead portmapper.
490 * - Some lockd implementations (e.g. HP) don't react to
491 * RPC_GRANTED calls; they seem to insist on RPC_GRANTED_MSG calls.
494 nlmsvc_grant_blocked(struct nlm_block *block)
496 struct nlm_file *file = block->b_file;
497 struct nlm_lock *lock = &block->b_call.a_args.lock;
498 struct file_lock *conflock;
501 dprintk("lockd: grant blocked lock %p\n", block);
503 /* First thing is lock the file */
506 /* Unlink block request from list */
507 nlmsvc_remove_block(block);
509 /* If b_granted is true this means we've been here before.
510 * Just retry the grant callback, possibly refreshing the RPC
512 if (block->b_granted) {
513 nlm_rebind_host(block->b_host);
517 /* Try the lock operation again */
518 if ((conflock = posix_test_lock(&file->f_file, &lock->fl)) != NULL) {
519 /* Bummer, we blocked again */
520 dprintk("lockd: lock still blocked\n");
521 nlmsvc_insert_block(block, NLM_NEVER);
522 posix_block_lock(conflock, &lock->fl);
527 /* Alright, no conflicting lock. Now lock it for real. If the
528 * following yields an error, this is most probably due to low
529 * memory. Retry the lock in a few seconds.
531 if ((error = posix_lock_file(&file->f_file, &lock->fl)) < 0) {
532 printk(KERN_WARNING "lockd: unexpected error %d in %s!\n",
533 -error, __FUNCTION__);
534 nlmsvc_insert_block(block, 10 * HZ);
540 /* Lock was granted by VFS. */
541 dprintk("lockd: GRANTing blocked lock.\n");
542 block->b_granted = 1;
545 /* Schedule next grant callback in 30 seconds */
546 nlmsvc_insert_block(block, 30 * HZ);
548 /* Call the client */
549 nlm_get_host(block->b_call.a_host);
550 if (nlmsvc_async_call(&block->b_call, NLMPROC_GRANTED_MSG,
551 nlmsvc_grant_callback) < 0)
552 nlm_release_host(block->b_call.a_host);
557 * This is the callback from the RPC layer when the NLM_GRANTED_MSG
558 * RPC call has succeeded or timed out.
559 * Like all RPC callbacks, it is invoked by the rpciod process, so it
560 * better not sleep. Therefore, we put the blocked lock on the nlm_blocked
561 * chain once more in order to have it removed by lockd itself (which can
562 * then sleep on the file semaphore without disrupting e.g. the nfs client).
565 nlmsvc_grant_callback(struct rpc_task *task)
567 struct nlm_rqst *call = (struct nlm_rqst *) task->tk_calldata;
568 struct nlm_block *block;
569 unsigned long timeout;
570 struct sockaddr_in *peer_addr = RPC_PEERADDR(task->tk_client);
572 dprintk("lockd: GRANT_MSG RPC callback\n");
573 dprintk("callback: looking for cookie %x, host (%08x)\n",
574 *(unsigned int *)(call->a_args.cookie.data),
575 ntohl(peer_addr->sin_addr.s_addr));
576 if (!(block = nlmsvc_find_block(&call->a_args.cookie, peer_addr))) {
577 dprintk("lockd: no block for cookie %x, host (%08x)\n",
578 *(u32 *)(call->a_args.cookie.data),
579 ntohl(peer_addr->sin_addr.s_addr));
583 /* Technically, we should down the file semaphore here. Since we
584 * move the block towards the head of the queue only, no harm
585 * can be done, though. */
586 if (task->tk_status < 0) {
587 /* RPC error: Re-insert for retransmission */
589 } else if (block->b_done) {
590 /* Block already removed, kill it for real */
593 /* Call was successful, now wait for client callback */
596 nlmsvc_insert_block(block, timeout);
597 svc_wake_up(block->b_daemon);
600 nlm_release_host(call->a_host);
604 * We received a GRANT_RES callback. Try to find the corresponding
608 nlmsvc_grant_reply(struct svc_rqst *rqstp, struct nlm_cookie *cookie, u32 status)
610 struct nlm_block *block;
611 struct nlm_file *file;
613 dprintk("grant_reply: looking for cookie %x, host (%08x), s=%d \n",
614 *(unsigned int *)(cookie->data),
615 ntohl(rqstp->rq_addr.sin_addr.s_addr), status);
616 if (!(block = nlmsvc_find_block(cookie, &rqstp->rq_addr)))
618 file = block->b_file;
622 if ((block = nlmsvc_find_block(cookie,&rqstp->rq_addr)) != NULL) {
623 if (status == NLM_LCK_DENIED_GRACE_PERIOD) {
624 /* Try again in a couple of seconds */
625 nlmsvc_insert_block(block, 10 * HZ);
628 /* Lock is now held by client, or has been rejected.
629 * In both cases, the block should be removed. */
632 if (status == NLM_LCK_GRANTED)
633 nlmsvc_delete_block(block, 0);
635 nlmsvc_delete_block(block, 1);
640 nlm_release_file(file);
644 * Retry all blocked locks that have been notified. This is where lockd
645 * picks up locks that can be granted, or grant notifications that must
649 nlmsvc_retry_blocked(void)
651 struct nlm_block *block;
653 dprintk("nlmsvc_retry_blocked(%p, when=%ld)\n",
655 nlm_blocked? nlm_blocked->b_when : 0);
656 while ((block = nlm_blocked)) {
657 if (block->b_when == NLM_NEVER)
659 if (time_after(block->b_when,jiffies))
661 dprintk("nlmsvc_retry_blocked(%p, when=%ld, done=%d)\n",
662 block, block->b_when, block->b_done);
664 nlmsvc_delete_block(block, 0);
666 nlmsvc_grant_blocked(block);
669 if ((block = nlm_blocked) && block->b_when != NLM_NEVER)
670 return (block->b_when - jiffies);
672 return MAX_SCHEDULE_TIMEOUT;