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 const struct rpc_call_ops nlmsvc_grant_ops;
49 * The list of blocked locks to retry
51 static struct nlm_block * nlm_blocked;
54 * Insert a blocked lock into the global list
57 nlmsvc_insert_block(struct nlm_block *block, unsigned long when)
59 struct nlm_block **bp, *b;
61 dprintk("lockd: nlmsvc_insert_block(%p, %ld)\n", block, when);
62 kref_get(&block->b_count);
64 nlmsvc_remove_block(block);
66 if (when != NLM_NEVER) {
67 if ((when += jiffies) == NLM_NEVER)
69 while ((b = *bp) && time_before_eq(b->b_when,when) && b->b_when != NLM_NEVER)
72 while ((b = *bp) != 0)
82 * Remove a block from the global list
85 nlmsvc_remove_block(struct nlm_block *block)
87 struct nlm_block **bp, *b;
91 for (bp = &nlm_blocked; (b = *bp) != 0; bp = &b->b_next) {
95 nlmsvc_release_block(block);
104 * Find a block for a given lock and optionally remove it from
107 static struct nlm_block *
108 nlmsvc_lookup_block(struct nlm_file *file, struct nlm_lock *lock, int remove)
110 struct nlm_block **head, *block;
111 struct file_lock *fl;
113 dprintk("lockd: nlmsvc_lookup_block f=%p pd=%d %Ld-%Ld ty=%d\n",
114 file, lock->fl.fl_pid,
115 (long long)lock->fl.fl_start,
116 (long long)lock->fl.fl_end, lock->fl.fl_type);
117 for (head = &nlm_blocked; (block = *head) != 0; head = &block->b_next) {
118 fl = &block->b_call.a_args.lock.fl;
119 dprintk("lockd: check f=%p pd=%d %Ld-%Ld ty=%d cookie=%s\n",
120 block->b_file, fl->fl_pid,
121 (long long)fl->fl_start,
122 (long long)fl->fl_end, fl->fl_type,
123 nlmdbg_cookie2a(&block->b_call.a_args.cookie));
124 if (block->b_file == file && nlm_compare_locks(fl, &lock->fl)) {
126 *head = block->b_next;
129 kref_get(&block->b_count);
137 static inline int nlm_cookie_match(struct nlm_cookie *a, struct nlm_cookie *b)
141 if(memcmp(a->data,b->data,a->len))
147 * Find a block with a given NLM cookie.
149 static inline struct nlm_block *
150 nlmsvc_find_block(struct nlm_cookie *cookie, struct sockaddr_in *sin)
152 struct nlm_block *block;
154 for (block = nlm_blocked; block; block = block->b_next) {
155 dprintk("cookie: head of blocked queue %p, block %p\n",
157 if (nlm_cookie_match(&block->b_call.a_args.cookie,cookie)
158 && nlm_cmp_addr(sin, &block->b_host->h_addr))
163 kref_get(&block->b_count);
168 * Create a block and initialize it.
170 * Note: we explicitly set the cookie of the grant reply to that of
171 * the blocked lock request. The spec explicitly mentions that the client
172 * should _not_ rely on the callback containing the same cookie as the
173 * request, but (as I found out later) that's because some implementations
174 * do just this. Never mind the standards comittees, they support our
175 * logging industries.
177 static inline struct nlm_block *
178 nlmsvc_create_block(struct svc_rqst *rqstp, struct nlm_file *file,
179 struct nlm_lock *lock, struct nlm_cookie *cookie)
181 struct nlm_block *block;
182 struct nlm_host *host;
183 struct nlm_rqst *call;
185 /* Create host handle for callback */
186 host = nlmclnt_lookup_host(&rqstp->rq_addr,
187 rqstp->rq_prot, rqstp->rq_vers);
191 /* Allocate memory for block, and initialize arguments */
192 if (!(block = (struct nlm_block *) kmalloc(sizeof(*block), GFP_KERNEL)))
194 memset(block, 0, sizeof(*block));
195 locks_init_lock(&block->b_call.a_args.lock.fl);
196 locks_init_lock(&block->b_call.a_res.lock.fl);
197 kref_init(&block->b_count);
199 if (!nlmclnt_setgrantargs(&block->b_call, lock))
202 /* Set notifier function for VFS, and init args */
203 block->b_call.a_args.lock.fl.fl_flags |= FL_SLEEP;
204 block->b_call.a_args.lock.fl.fl_lmops = &nlmsvc_lock_operations;
205 block->b_call.a_args.cookie = *cookie; /* see above */
207 dprintk("lockd: created block %p...\n", block);
209 /* Create and initialize the block */
210 block->b_daemon = rqstp->rq_server;
211 block->b_host = host;
212 block->b_file = file;
214 /* Add to file's list of blocks */
215 block->b_fnext = file->f_blocks;
216 file->f_blocks = block;
218 /* Set up RPC arguments for callback */
219 call = &block->b_call;
221 call->a_flags = RPC_TASK_ASYNC;
228 nlm_release_host(host);
233 * Delete a block. If the lock was cancelled or the grant callback
234 * failed, unlock is set to 1.
235 * It is the caller's responsibility to check whether the file
236 * can be closed hereafter.
238 static int nlmsvc_unlink_block(struct nlm_block *block)
241 dprintk("lockd: unlinking block %p...\n", block);
243 /* Remove block from list */
244 status = posix_unblock_lock(block->b_file->f_file, &block->b_call.a_args.lock.fl);
245 nlmsvc_remove_block(block);
249 static void nlmsvc_free_block(struct kref *kref)
251 struct nlm_block *block = container_of(kref, struct nlm_block, b_count);
252 struct nlm_file *file = block->b_file;
253 struct nlm_block **bp;
255 dprintk("lockd: freeing block %p...\n", block);
257 /* Remove block from file's list of blocks */
258 for (bp = &file->f_blocks; *bp; bp = &(*bp)->b_fnext) {
260 *bp = block->b_fnext;
266 nlm_release_host(block->b_host);
267 nlmclnt_freegrantargs(&block->b_call);
271 static void nlmsvc_release_block(struct nlm_block *block)
274 kref_put(&block->b_count, nlmsvc_free_block);
278 * Loop over all blocks and perform the action specified.
279 * (NLM_ACT_CHECK handled by nlmsvc_inspect_file).
282 nlmsvc_traverse_blocks(struct nlm_host *host, struct nlm_file *file, int action)
284 struct nlm_block *block, *next;
285 /* XXX: Will everything get cleaned up if we don't unlock here? */
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_unlink_block(block);
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 nlm_block *block, *newblock = NULL;
313 dprintk("lockd: nlmsvc_lock(%s/%ld, ty=%d, pi=%d, %Ld-%Ld, bl=%d)\n",
314 file->f_file->f_dentry->d_inode->i_sb->s_id,
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 lock->fl.fl_flags &= ~FL_SLEEP;
324 /* Lock file against concurrent access */
326 /* Get existing block (in case client is busy-waiting) */
327 block = nlmsvc_lookup_block(file, lock, 0);
329 if (newblock != NULL)
330 lock = &newblock->b_call.a_args.lock;
332 lock = &block->b_call.a_args.lock;
334 error = posix_lock_file(file->f_file, &lock->fl);
335 lock->fl.fl_flags &= ~FL_SLEEP;
337 dprintk("lockd: posix_lock_file returned %d\n", error);
348 default: /* includes ENOLCK */
349 ret = nlm_lck_denied_nolocks;
353 ret = nlm_lck_denied;
357 ret = nlm_lck_blocked;
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 * to claim it while doing host garbage collection */
365 if (newblock == NULL) {
367 dprintk("lockd: blocking on this lock (allocating).\n");
368 if (!(newblock = nlmsvc_create_block(rqstp, file, lock, cookie)))
369 return nlm_lck_denied_nolocks;
373 /* Append to list of blocked */
374 nlmsvc_insert_block(newblock, NLM_NEVER);
377 nlmsvc_release_block(newblock);
378 nlmsvc_release_block(block);
379 dprintk("lockd: nlmsvc_lock returned %u\n", ret);
384 * Test for presence of a conflicting lock.
387 nlmsvc_testlock(struct nlm_file *file, struct nlm_lock *lock,
388 struct nlm_lock *conflock)
390 dprintk("lockd: nlmsvc_testlock(%s/%ld, ty=%d, %Ld-%Ld)\n",
391 file->f_file->f_dentry->d_inode->i_sb->s_id,
392 file->f_file->f_dentry->d_inode->i_ino,
394 (long long)lock->fl.fl_start,
395 (long long)lock->fl.fl_end);
397 if (posix_test_lock(file->f_file, &lock->fl, &conflock->fl)) {
398 dprintk("lockd: conflicting lock(ty=%d, %Ld-%Ld)\n",
399 conflock->fl.fl_type,
400 (long long)conflock->fl.fl_start,
401 (long long)conflock->fl.fl_end);
402 conflock->caller = "somehost"; /* FIXME */
403 conflock->oh.len = 0; /* don't return OH info */
404 conflock->svid = conflock->fl.fl_pid;
405 return nlm_lck_denied;
413 * This implies a CANCEL call: We send a GRANT_MSG, the client replies
414 * with a GRANT_RES call which gets lost, and calls UNLOCK immediately
415 * afterwards. In this case the block will still be there, and hence
419 nlmsvc_unlock(struct nlm_file *file, struct nlm_lock *lock)
423 dprintk("lockd: nlmsvc_unlock(%s/%ld, pi=%d, %Ld-%Ld)\n",
424 file->f_file->f_dentry->d_inode->i_sb->s_id,
425 file->f_file->f_dentry->d_inode->i_ino,
427 (long long)lock->fl.fl_start,
428 (long long)lock->fl.fl_end);
430 /* First, cancel any lock that might be there */
431 nlmsvc_cancel_blocked(file, lock);
433 lock->fl.fl_type = F_UNLCK;
434 error = posix_lock_file(file->f_file, &lock->fl);
436 return (error < 0)? nlm_lck_denied_nolocks : nlm_granted;
440 * Cancel a previously blocked request.
442 * A cancel request always overrides any grant that may currently
444 * The calling procedure must check whether the file can be closed.
447 nlmsvc_cancel_blocked(struct nlm_file *file, struct nlm_lock *lock)
449 struct nlm_block *block;
452 dprintk("lockd: nlmsvc_cancel(%s/%ld, pi=%d, %Ld-%Ld)\n",
453 file->f_file->f_dentry->d_inode->i_sb->s_id,
454 file->f_file->f_dentry->d_inode->i_ino,
456 (long long)lock->fl.fl_start,
457 (long long)lock->fl.fl_end);
460 if ((block = nlmsvc_lookup_block(file, lock, 1)) != NULL) {
461 status = nlmsvc_unlink_block(block);
462 nlmsvc_release_block(block);
465 return status ? nlm_lck_denied : nlm_granted;
469 * Unblock a blocked lock request. This is a callback invoked from the
470 * VFS layer when a lock on which we blocked is removed.
472 * This function doesn't grant the blocked lock instantly, but rather moves
473 * the block to the head of nlm_blocked where it can be picked up by lockd.
476 nlmsvc_notify_blocked(struct file_lock *fl)
478 struct nlm_block **bp, *block;
480 dprintk("lockd: VFS unblock notification for block %p\n", fl);
481 for (bp = &nlm_blocked; (block = *bp) != 0; bp = &block->b_next) {
482 if (nlm_compare_locks(&block->b_call.a_args.lock.fl, fl)) {
483 nlmsvc_insert_block(block, 0);
484 svc_wake_up(block->b_daemon);
489 printk(KERN_WARNING "lockd: notification for unknown block!\n");
492 static int nlmsvc_same_owner(struct file_lock *fl1, struct file_lock *fl2)
494 return fl1->fl_owner == fl2->fl_owner && fl1->fl_pid == fl2->fl_pid;
497 struct lock_manager_operations nlmsvc_lock_operations = {
498 .fl_compare_owner = nlmsvc_same_owner,
499 .fl_notify = nlmsvc_notify_blocked,
503 * Try to claim a lock that was previously blocked.
505 * Note that we use both the RPC_GRANTED_MSG call _and_ an async
506 * RPC thread when notifying the client. This seems like overkill...
508 * - we don't want to use a synchronous RPC thread, otherwise
509 * we might find ourselves hanging on a dead portmapper.
510 * - Some lockd implementations (e.g. HP) don't react to
511 * RPC_GRANTED calls; they seem to insist on RPC_GRANTED_MSG calls.
514 nlmsvc_grant_blocked(struct nlm_block *block)
516 struct nlm_file *file = block->b_file;
517 struct nlm_lock *lock = &block->b_call.a_args.lock;
520 dprintk("lockd: grant blocked lock %p\n", block);
522 /* First thing is lock the file */
525 /* Unlink block request from list */
526 nlmsvc_unlink_block(block);
528 /* If b_granted is true this means we've been here before.
529 * Just retry the grant callback, possibly refreshing the RPC
531 if (block->b_granted) {
532 nlm_rebind_host(block->b_host);
536 /* Try the lock operation again */
537 lock->fl.fl_flags |= FL_SLEEP;
538 error = posix_lock_file(file->f_file, &lock->fl);
539 lock->fl.fl_flags &= ~FL_SLEEP;
545 dprintk("lockd: lock still blocked\n");
546 nlmsvc_insert_block(block, NLM_NEVER);
549 printk(KERN_WARNING "lockd: unexpected error %d in %s!\n",
550 -error, __FUNCTION__);
551 nlmsvc_insert_block(block, 10 * HZ);
556 /* Lock was granted by VFS. */
557 dprintk("lockd: GRANTing blocked lock.\n");
558 block->b_granted = 1;
560 /* Schedule next grant callback in 30 seconds */
561 nlmsvc_insert_block(block, 30 * HZ);
563 /* Call the client */
564 kref_get(&block->b_count);
565 if (nlmsvc_async_call(&block->b_call, NLMPROC_GRANTED_MSG,
566 &nlmsvc_grant_ops) < 0)
567 nlmsvc_release_block(block);
573 * This is the callback from the RPC layer when the NLM_GRANTED_MSG
574 * RPC call has succeeded or timed out.
575 * Like all RPC callbacks, it is invoked by the rpciod process, so it
576 * better not sleep. Therefore, we put the blocked lock on the nlm_blocked
577 * chain once more in order to have it removed by lockd itself (which can
578 * then sleep on the file semaphore without disrupting e.g. the nfs client).
580 static void nlmsvc_grant_callback(struct rpc_task *task, void *data)
582 struct nlm_rqst *call = data;
583 struct nlm_block *block = container_of(call, struct nlm_block, b_call);
584 unsigned long timeout;
586 dprintk("lockd: GRANT_MSG RPC callback\n");
588 /* Technically, we should down the file semaphore here. Since we
589 * move the block towards the head of the queue only, no harm
590 * can be done, though. */
591 if (task->tk_status < 0) {
592 /* RPC error: Re-insert for retransmission */
594 } else if (block->b_done) {
595 /* Block already removed, kill it for real */
598 /* Call was successful, now wait for client callback */
601 nlmsvc_insert_block(block, timeout);
602 svc_wake_up(block->b_daemon);
603 nlmsvc_release_block(block);
606 static const struct rpc_call_ops nlmsvc_grant_ops = {
607 .rpc_call_done = nlmsvc_grant_callback,
611 * We received a GRANT_RES callback. Try to find the corresponding
615 nlmsvc_grant_reply(struct svc_rqst *rqstp, struct nlm_cookie *cookie, u32 status)
617 struct nlm_block *block;
618 struct nlm_file *file;
620 dprintk("grant_reply: looking for cookie %x, host (%08x), s=%d \n",
621 *(unsigned int *)(cookie->data),
622 ntohl(rqstp->rq_addr.sin_addr.s_addr), status);
623 if (!(block = nlmsvc_find_block(cookie, &rqstp->rq_addr)))
625 file = block->b_file;
630 if (status == NLM_LCK_DENIED_GRACE_PERIOD) {
631 /* Try again in a couple of seconds */
632 nlmsvc_insert_block(block, 10 * HZ);
634 /* Lock is now held by client, or has been rejected.
635 * In both cases, the block should be removed. */
636 nlmsvc_unlink_block(block);
640 nlm_release_file(file);
641 nlmsvc_release_block(block);
645 * Retry all blocked locks that have been notified. This is where lockd
646 * picks up locks that can be granted, or grant notifications that must
650 nlmsvc_retry_blocked(void)
652 struct nlm_block *block;
654 dprintk("nlmsvc_retry_blocked(%p, when=%ld)\n",
656 nlm_blocked? nlm_blocked->b_when : 0);
657 while ((block = nlm_blocked) != 0) {
658 if (block->b_when == NLM_NEVER)
660 if (time_after(block->b_when,jiffies))
662 dprintk("nlmsvc_retry_blocked(%p, when=%ld, done=%d)\n",
663 block, block->b_when, block->b_done);
664 kref_get(&block->b_count);
666 nlmsvc_unlink_block(block);
668 nlmsvc_grant_blocked(block);
669 nlmsvc_release_block(block);
672 if ((block = nlm_blocked) && block->b_when != NLM_NEVER)
673 return (block->b_when - jiffies);
675 return MAX_SCHEDULE_TIMEOUT;