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/types.h>
24 #include <linux/errno.h>
25 #include <linux/kernel.h>
26 #include <linux/sched.h>
27 #include <linux/smp_lock.h>
28 #include <linux/sunrpc/clnt.h>
29 #include <linux/sunrpc/svc.h>
30 #include <linux/lockd/nlm.h>
31 #include <linux/lockd/lockd.h>
33 #define NLMDBG_FACILITY NLMDBG_SVCLOCK
35 #ifdef CONFIG_LOCKD_V4
36 #define nlm_deadlock nlm4_deadlock
38 #define nlm_deadlock nlm_lck_denied
41 static void nlmsvc_release_block(struct nlm_block
*block
);
42 static void nlmsvc_insert_block(struct nlm_block
*block
, unsigned long);
43 static void nlmsvc_remove_block(struct nlm_block
*block
);
45 static int nlmsvc_setgrantargs(struct nlm_rqst
*call
, struct nlm_lock
*lock
);
46 static void nlmsvc_freegrantargs(struct nlm_rqst
*call
);
47 static const struct rpc_call_ops nlmsvc_grant_ops
;
50 * The list of blocked locks to retry
52 static LIST_HEAD(nlm_blocked
);
55 * Insert a blocked lock into the global list
58 nlmsvc_insert_block(struct nlm_block
*block
, unsigned long when
)
61 struct list_head
*pos
;
63 dprintk("lockd: nlmsvc_insert_block(%p, %ld)\n", block
, when
);
64 if (list_empty(&block
->b_list
)) {
65 kref_get(&block
->b_count
);
67 list_del_init(&block
->b_list
);
71 if (when
!= NLM_NEVER
) {
72 if ((when
+= jiffies
) == NLM_NEVER
)
74 list_for_each(pos
, &nlm_blocked
) {
75 b
= list_entry(pos
, struct nlm_block
, b_list
);
76 if (time_after(b
->b_when
,when
) || b
->b_when
== NLM_NEVER
)
79 /* On normal exit from the loop, pos == &nlm_blocked,
80 * so we will be adding to the end of the list - good
84 list_add_tail(&block
->b_list
, pos
);
89 * Remove a block from the global list
92 nlmsvc_remove_block(struct nlm_block
*block
)
94 if (!list_empty(&block
->b_list
)) {
95 list_del_init(&block
->b_list
);
96 nlmsvc_release_block(block
);
101 * Find a block for a given lock
103 static struct nlm_block
*
104 nlmsvc_lookup_block(struct nlm_file
*file
, struct nlm_lock
*lock
)
106 struct nlm_block
*block
;
107 struct file_lock
*fl
;
109 dprintk("lockd: nlmsvc_lookup_block f=%p pd=%d %Ld-%Ld ty=%d\n",
110 file
, lock
->fl
.fl_pid
,
111 (long long)lock
->fl
.fl_start
,
112 (long long)lock
->fl
.fl_end
, lock
->fl
.fl_type
);
113 list_for_each_entry(block
, &nlm_blocked
, b_list
) {
114 fl
= &block
->b_call
->a_args
.lock
.fl
;
115 dprintk("lockd: check f=%p pd=%d %Ld-%Ld ty=%d cookie=%s\n",
116 block
->b_file
, fl
->fl_pid
,
117 (long long)fl
->fl_start
,
118 (long long)fl
->fl_end
, fl
->fl_type
,
119 nlmdbg_cookie2a(&block
->b_call
->a_args
.cookie
));
120 if (block
->b_file
== file
&& nlm_compare_locks(fl
, &lock
->fl
)) {
121 kref_get(&block
->b_count
);
129 static inline int nlm_cookie_match(struct nlm_cookie
*a
, struct nlm_cookie
*b
)
133 if(memcmp(a
->data
,b
->data
,a
->len
))
139 * Find a block with a given NLM cookie.
141 static inline struct nlm_block
*
142 nlmsvc_find_block(struct nlm_cookie
*cookie
)
144 struct nlm_block
*block
;
146 list_for_each_entry(block
, &nlm_blocked
, b_list
) {
147 if (nlm_cookie_match(&block
->b_call
->a_args
.cookie
,cookie
))
154 dprintk("nlmsvc_find_block(%s): block=%p\n", nlmdbg_cookie2a(cookie
), block
);
155 kref_get(&block
->b_count
);
160 * Create a block and initialize it.
162 * Note: we explicitly set the cookie of the grant reply to that of
163 * the blocked lock request. The spec explicitly mentions that the client
164 * should _not_ rely on the callback containing the same cookie as the
165 * request, but (as I found out later) that's because some implementations
166 * do just this. Never mind the standards comittees, they support our
167 * logging industries.
169 * 10 years later: I hope we can safely ignore these old and broken
170 * clients by now. Let's fix this so we can uniquely identify an incoming
171 * GRANTED_RES message by cookie, without having to rely on the client's IP
174 static inline struct nlm_block
*
175 nlmsvc_create_block(struct svc_rqst
*rqstp
, struct nlm_file
*file
,
176 struct nlm_lock
*lock
, struct nlm_cookie
*cookie
)
178 struct nlm_block
*block
;
179 struct nlm_host
*host
;
180 struct nlm_rqst
*call
= NULL
;
182 /* Create host handle for callback */
183 host
= nlmsvc_lookup_host(rqstp
, lock
->caller
, lock
->len
);
187 call
= nlm_alloc_call(host
);
191 /* Allocate memory for block, and initialize arguments */
192 block
= kzalloc(sizeof(*block
), GFP_KERNEL
);
195 kref_init(&block
->b_count
);
196 INIT_LIST_HEAD(&block
->b_list
);
197 INIT_LIST_HEAD(&block
->b_flist
);
199 if (!nlmsvc_setgrantargs(call
, lock
))
202 /* Set notifier function for VFS, and init args */
203 call
->a_args
.lock
.fl
.fl_flags
|= FL_SLEEP
;
204 call
->a_args
.lock
.fl
.fl_lmops
= &nlmsvc_lock_operations
;
205 nlmclnt_next_cookie(&call
->a_args
.cookie
);
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
;
215 /* Add to file's list of blocks */
216 list_add(&block
->b_flist
, &file
->f_blocks
);
218 /* Set up RPC arguments for callback */
219 block
->b_call
= call
;
220 call
->a_flags
= RPC_TASK_ASYNC
;
221 call
->a_block
= block
;
228 nlm_release_call(call
);
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
;
254 dprintk("lockd: freeing block %p...\n", block
);
256 /* Remove block from file's list of blocks */
258 list_del_init(&block
->b_flist
);
261 nlmsvc_freegrantargs(block
->b_call
);
262 nlm_release_call(block
->b_call
);
263 nlm_release_file(block
->b_file
);
267 static void nlmsvc_release_block(struct nlm_block
*block
)
270 kref_put(&block
->b_count
, nlmsvc_free_block
);
274 * Loop over all blocks and delete blocks held by
277 void nlmsvc_traverse_blocks(struct nlm_host
*host
,
278 struct nlm_file
*file
,
279 nlm_host_match_fn_t match
)
281 struct nlm_block
*block
, *next
;
285 list_for_each_entry_safe(block
, next
, &file
->f_blocks
, b_flist
) {
286 if (!match(block
->b_host
, host
))
288 /* Do not destroy blocks that are not on
289 * the global retry list - why? */
290 if (list_empty(&block
->b_list
))
292 kref_get(&block
->b_count
);
294 nlmsvc_unlink_block(block
);
295 nlmsvc_release_block(block
);
302 * Initialize arguments for GRANTED call. The nlm_rqst structure
303 * has been cleared already.
305 static int nlmsvc_setgrantargs(struct nlm_rqst
*call
, struct nlm_lock
*lock
)
307 locks_copy_lock(&call
->a_args
.lock
.fl
, &lock
->fl
);
308 memcpy(&call
->a_args
.lock
.fh
, &lock
->fh
, sizeof(call
->a_args
.lock
.fh
));
309 call
->a_args
.lock
.caller
= utsname()->nodename
;
310 call
->a_args
.lock
.oh
.len
= lock
->oh
.len
;
312 /* set default data area */
313 call
->a_args
.lock
.oh
.data
= call
->a_owner
;
314 call
->a_args
.lock
.svid
= lock
->fl
.fl_pid
;
316 if (lock
->oh
.len
> NLMCLNT_OHSIZE
) {
317 void *data
= kmalloc(lock
->oh
.len
, GFP_KERNEL
);
320 call
->a_args
.lock
.oh
.data
= (u8
*) data
;
323 memcpy(call
->a_args
.lock
.oh
.data
, lock
->oh
.data
, lock
->oh
.len
);
327 static void nlmsvc_freegrantargs(struct nlm_rqst
*call
)
329 if (call
->a_args
.lock
.oh
.data
!= call
->a_owner
)
330 kfree(call
->a_args
.lock
.oh
.data
);
334 * Attempt to establish a lock, and if it can't be granted, block it
338 nlmsvc_lock(struct svc_rqst
*rqstp
, struct nlm_file
*file
,
339 struct nlm_lock
*lock
, int wait
, struct nlm_cookie
*cookie
)
341 struct nlm_block
*block
, *newblock
= NULL
;
345 dprintk("lockd: nlmsvc_lock(%s/%ld, ty=%d, pi=%d, %Ld-%Ld, bl=%d)\n",
346 file
->f_file
->f_dentry
->d_inode
->i_sb
->s_id
,
347 file
->f_file
->f_dentry
->d_inode
->i_ino
,
348 lock
->fl
.fl_type
, lock
->fl
.fl_pid
,
349 (long long)lock
->fl
.fl_start
,
350 (long long)lock
->fl
.fl_end
,
354 lock
->fl
.fl_flags
&= ~FL_SLEEP
;
356 /* Lock file against concurrent access */
358 /* Get existing block (in case client is busy-waiting) */
359 block
= nlmsvc_lookup_block(file
, lock
);
361 if (newblock
!= NULL
)
362 lock
= &newblock
->b_call
->a_args
.lock
;
364 lock
= &block
->b_call
->a_args
.lock
;
366 error
= posix_lock_file(file
->f_file
, &lock
->fl
);
367 lock
->fl
.fl_flags
&= ~FL_SLEEP
;
369 dprintk("lockd: posix_lock_file returned %d\n", error
);
380 default: /* includes ENOLCK */
381 ret
= nlm_lck_denied_nolocks
;
385 ret
= nlm_lck_denied
;
389 ret
= nlm_lck_blocked
;
393 /* If we don't have a block, create and initialize it. Then
394 * retry because we may have slept in kmalloc. */
395 /* We have to release f_sema as nlmsvc_create_block may try to
396 * to claim it while doing host garbage collection */
397 if (newblock
== NULL
) {
399 dprintk("lockd: blocking on this lock (allocating).\n");
400 if (!(newblock
= nlmsvc_create_block(rqstp
, file
, lock
, cookie
)))
401 return nlm_lck_denied_nolocks
;
405 /* Append to list of blocked */
406 nlmsvc_insert_block(newblock
, NLM_NEVER
);
409 nlmsvc_release_block(newblock
);
410 nlmsvc_release_block(block
);
411 dprintk("lockd: nlmsvc_lock returned %u\n", ret
);
416 * Test for presence of a conflicting lock.
419 nlmsvc_testlock(struct nlm_file
*file
, struct nlm_lock
*lock
,
420 struct nlm_lock
*conflock
)
422 dprintk("lockd: nlmsvc_testlock(%s/%ld, ty=%d, %Ld-%Ld)\n",
423 file
->f_file
->f_dentry
->d_inode
->i_sb
->s_id
,
424 file
->f_file
->f_dentry
->d_inode
->i_ino
,
426 (long long)lock
->fl
.fl_start
,
427 (long long)lock
->fl
.fl_end
);
429 if (posix_test_lock(file
->f_file
, &lock
->fl
, &conflock
->fl
)) {
430 dprintk("lockd: conflicting lock(ty=%d, %Ld-%Ld)\n",
431 conflock
->fl
.fl_type
,
432 (long long)conflock
->fl
.fl_start
,
433 (long long)conflock
->fl
.fl_end
);
434 conflock
->caller
= "somehost"; /* FIXME */
435 conflock
->len
= strlen(conflock
->caller
);
436 conflock
->oh
.len
= 0; /* don't return OH info */
437 conflock
->svid
= conflock
->fl
.fl_pid
;
438 return nlm_lck_denied
;
446 * This implies a CANCEL call: We send a GRANT_MSG, the client replies
447 * with a GRANT_RES call which gets lost, and calls UNLOCK immediately
448 * afterwards. In this case the block will still be there, and hence
452 nlmsvc_unlock(struct nlm_file
*file
, struct nlm_lock
*lock
)
456 dprintk("lockd: nlmsvc_unlock(%s/%ld, pi=%d, %Ld-%Ld)\n",
457 file
->f_file
->f_dentry
->d_inode
->i_sb
->s_id
,
458 file
->f_file
->f_dentry
->d_inode
->i_ino
,
460 (long long)lock
->fl
.fl_start
,
461 (long long)lock
->fl
.fl_end
);
463 /* First, cancel any lock that might be there */
464 nlmsvc_cancel_blocked(file
, lock
);
466 lock
->fl
.fl_type
= F_UNLCK
;
467 error
= posix_lock_file(file
->f_file
, &lock
->fl
);
469 return (error
< 0)? nlm_lck_denied_nolocks
: nlm_granted
;
473 * Cancel a previously blocked request.
475 * A cancel request always overrides any grant that may currently
477 * The calling procedure must check whether the file can be closed.
480 nlmsvc_cancel_blocked(struct nlm_file
*file
, struct nlm_lock
*lock
)
482 struct nlm_block
*block
;
485 dprintk("lockd: nlmsvc_cancel(%s/%ld, pi=%d, %Ld-%Ld)\n",
486 file
->f_file
->f_dentry
->d_inode
->i_sb
->s_id
,
487 file
->f_file
->f_dentry
->d_inode
->i_ino
,
489 (long long)lock
->fl
.fl_start
,
490 (long long)lock
->fl
.fl_end
);
493 block
= nlmsvc_lookup_block(file
, lock
);
496 status
= nlmsvc_unlink_block(block
);
497 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
*block
;
514 dprintk("lockd: VFS unblock notification for block %p\n", fl
);
515 list_for_each_entry(block
, &nlm_blocked
, b_list
) {
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 /* Unlink block request from list */
557 nlmsvc_unlink_block(block
);
559 /* If b_granted is true this means we've been here before.
560 * Just retry the grant callback, possibly refreshing the RPC
562 if (block
->b_granted
) {
563 nlm_rebind_host(block
->b_host
);
567 /* Try the lock operation again */
568 lock
->fl
.fl_flags
|= FL_SLEEP
;
569 error
= posix_lock_file(file
->f_file
, &lock
->fl
);
570 lock
->fl
.fl_flags
&= ~FL_SLEEP
;
576 dprintk("lockd: lock still blocked\n");
577 nlmsvc_insert_block(block
, NLM_NEVER
);
580 printk(KERN_WARNING
"lockd: unexpected error %d in %s!\n",
581 -error
, __FUNCTION__
);
582 nlmsvc_insert_block(block
, 10 * HZ
);
587 /* Lock was granted by VFS. */
588 dprintk("lockd: GRANTing blocked lock.\n");
589 block
->b_granted
= 1;
591 /* Schedule next grant callback in 30 seconds */
592 nlmsvc_insert_block(block
, 30 * HZ
);
594 /* Call the client */
595 kref_get(&block
->b_count
);
596 if (nlm_async_call(block
->b_call
, NLMPROC_GRANTED_MSG
,
597 &nlmsvc_grant_ops
) < 0)
598 nlmsvc_release_block(block
);
602 * This is the callback from the RPC layer when the NLM_GRANTED_MSG
603 * RPC call has succeeded or timed out.
604 * Like all RPC callbacks, it is invoked by the rpciod process, so it
605 * better not sleep. Therefore, we put the blocked lock on the nlm_blocked
606 * chain once more in order to have it removed by lockd itself (which can
607 * then sleep on the file semaphore without disrupting e.g. the nfs client).
609 static void nlmsvc_grant_callback(struct rpc_task
*task
, void *data
)
611 struct nlm_rqst
*call
= data
;
612 struct nlm_block
*block
= call
->a_block
;
613 unsigned long timeout
;
615 dprintk("lockd: GRANT_MSG RPC callback\n");
617 /* Technically, we should down the file semaphore here. Since we
618 * move the block towards the head of the queue only, no harm
619 * can be done, though. */
620 if (task
->tk_status
< 0) {
621 /* RPC error: Re-insert for retransmission */
624 /* Call was successful, now wait for client callback */
627 nlmsvc_insert_block(block
, timeout
);
628 svc_wake_up(block
->b_daemon
);
631 static void nlmsvc_grant_release(void *data
)
633 struct nlm_rqst
*call
= data
;
635 nlmsvc_release_block(call
->a_block
);
638 static const struct rpc_call_ops nlmsvc_grant_ops
= {
639 .rpc_call_done
= nlmsvc_grant_callback
,
640 .rpc_release
= nlmsvc_grant_release
,
644 * We received a GRANT_RES callback. Try to find the corresponding
648 nlmsvc_grant_reply(struct nlm_cookie
*cookie
, u32 status
)
650 struct nlm_block
*block
;
652 dprintk("grant_reply: looking for cookie %x, s=%d \n",
653 *(unsigned int *)(cookie
->data
), status
);
654 if (!(block
= nlmsvc_find_block(cookie
)))
658 if (status
== NLM_LCK_DENIED_GRACE_PERIOD
) {
659 /* Try again in a couple of seconds */
660 nlmsvc_insert_block(block
, 10 * HZ
);
662 /* Lock is now held by client, or has been rejected.
663 * In both cases, the block should be removed. */
664 nlmsvc_unlink_block(block
);
667 nlmsvc_release_block(block
);
671 * Retry all blocked locks that have been notified. This is where lockd
672 * picks up locks that can be granted, or grant notifications that must
676 nlmsvc_retry_blocked(void)
678 unsigned long timeout
= MAX_SCHEDULE_TIMEOUT
;
679 struct nlm_block
*block
;
681 while (!list_empty(&nlm_blocked
)) {
682 block
= list_entry(nlm_blocked
.next
, struct nlm_block
, b_list
);
684 if (block
->b_when
== NLM_NEVER
)
686 if (time_after(block
->b_when
,jiffies
)) {
687 timeout
= block
->b_when
- jiffies
;
691 dprintk("nlmsvc_retry_blocked(%p, when=%ld)\n",
692 block
, block
->b_when
);
693 kref_get(&block
->b_count
);
694 nlmsvc_grant_blocked(block
);
695 nlmsvc_release_block(block
);