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
108 static struct nlm_block
*
109 nlmsvc_lookup_block(struct nlm_file
*file
, struct nlm_lock
*lock
)
111 struct nlm_block
**head
, *block
;
112 struct file_lock
*fl
;
114 dprintk("lockd: nlmsvc_lookup_block f=%p pd=%d %Ld-%Ld ty=%d\n",
115 file
, lock
->fl
.fl_pid
,
116 (long long)lock
->fl
.fl_start
,
117 (long long)lock
->fl
.fl_end
, lock
->fl
.fl_type
);
118 for (head
= &nlm_blocked
; (block
= *head
) != 0; head
= &block
->b_next
) {
119 fl
= &block
->b_call
->a_args
.lock
.fl
;
120 dprintk("lockd: check f=%p pd=%d %Ld-%Ld ty=%d cookie=%s\n",
121 block
->b_file
, fl
->fl_pid
,
122 (long long)fl
->fl_start
,
123 (long long)fl
->fl_end
, fl
->fl_type
,
124 nlmdbg_cookie2a(&block
->b_call
->a_args
.cookie
));
125 if (block
->b_file
== file
&& nlm_compare_locks(fl
, &lock
->fl
)) {
126 kref_get(&block
->b_count
);
134 static inline int nlm_cookie_match(struct nlm_cookie
*a
, struct nlm_cookie
*b
)
138 if(memcmp(a
->data
,b
->data
,a
->len
))
144 * Find a block with a given NLM cookie.
146 static inline struct nlm_block
*
147 nlmsvc_find_block(struct nlm_cookie
*cookie
, struct sockaddr_in
*sin
)
149 struct nlm_block
*block
;
151 for (block
= nlm_blocked
; block
; block
= block
->b_next
) {
152 dprintk("cookie: head of blocked queue %p, block %p\n",
154 if (nlm_cookie_match(&block
->b_call
->a_args
.cookie
,cookie
)
155 && nlm_cmp_addr(sin
, &block
->b_host
->h_addr
))
160 kref_get(&block
->b_count
);
165 * Create a block and initialize it.
167 * Note: we explicitly set the cookie of the grant reply to that of
168 * the blocked lock request. The spec explicitly mentions that the client
169 * should _not_ rely on the callback containing the same cookie as the
170 * request, but (as I found out later) that's because some implementations
171 * do just this. Never mind the standards comittees, they support our
172 * logging industries.
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
);
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
);
197 if (!nlmsvc_setgrantargs(call
, lock
))
200 /* Set notifier function for VFS, and init args */
201 call
->a_args
.lock
.fl
.fl_flags
|= FL_SLEEP
;
202 call
->a_args
.lock
.fl
.fl_lmops
= &nlmsvc_lock_operations
;
203 call
->a_args
.cookie
= *cookie
; /* see above */
205 dprintk("lockd: created block %p...\n", block
);
207 /* Create and initialize the block */
208 block
->b_daemon
= rqstp
->rq_server
;
209 block
->b_host
= host
;
210 block
->b_file
= file
;
213 /* Add to file's list of blocks */
214 block
->b_fnext
= file
->f_blocks
;
215 file
->f_blocks
= block
;
217 /* Set up RPC arguments for callback */
218 block
->b_call
= call
;
219 call
->a_flags
= RPC_TASK_ASYNC
;
220 call
->a_block
= block
;
227 nlm_release_call(call
);
232 * Delete a block. If the lock was cancelled or the grant callback
233 * failed, unlock is set to 1.
234 * It is the caller's responsibility to check whether the file
235 * can be closed hereafter.
237 static int nlmsvc_unlink_block(struct nlm_block
*block
)
240 dprintk("lockd: unlinking block %p...\n", block
);
242 /* Remove block from list */
243 status
= posix_unblock_lock(block
->b_file
->f_file
, &block
->b_call
->a_args
.lock
.fl
);
244 nlmsvc_remove_block(block
);
248 static void nlmsvc_free_block(struct kref
*kref
)
250 struct nlm_block
*block
= container_of(kref
, struct nlm_block
, b_count
);
251 struct nlm_file
*file
= block
->b_file
;
252 struct nlm_block
**bp
;
254 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 nlmsvc_freegrantargs(block
->b_call
);
267 nlm_release_call(block
->b_call
);
268 nlm_release_file(block
->b_file
);
272 static void nlmsvc_release_block(struct nlm_block
*block
)
275 kref_put(&block
->b_count
, nlmsvc_free_block
);
278 static void nlmsvc_act_mark(struct nlm_host
*host
, struct nlm_file
*file
)
280 struct nlm_block
*block
;
283 for (block
= file
->f_blocks
; block
!= NULL
; block
= block
->b_fnext
)
284 block
->b_host
->h_inuse
= 1;
288 static void nlmsvc_act_unlock(struct nlm_host
*host
, struct nlm_file
*file
)
290 struct nlm_block
*block
;
294 for (block
= file
->f_blocks
; block
!= NULL
; block
= block
->b_fnext
) {
295 if (host
!= NULL
&& host
!= block
->b_host
)
297 if (!block
->b_queued
)
299 kref_get(&block
->b_count
);
301 nlmsvc_unlink_block(block
);
302 nlmsvc_release_block(block
);
309 * Loop over all blocks and perform the action specified.
310 * (NLM_ACT_CHECK handled by nlmsvc_inspect_file).
313 nlmsvc_traverse_blocks(struct nlm_host
*host
, struct nlm_file
*file
, int action
)
315 if (action
== NLM_ACT_MARK
)
316 nlmsvc_act_mark(host
, file
);
318 nlmsvc_act_unlock(host
, file
);
322 * Initialize arguments for GRANTED call. The nlm_rqst structure
323 * has been cleared already.
325 static int nlmsvc_setgrantargs(struct nlm_rqst
*call
, struct nlm_lock
*lock
)
327 locks_copy_lock(&call
->a_args
.lock
.fl
, &lock
->fl
);
328 memcpy(&call
->a_args
.lock
.fh
, &lock
->fh
, sizeof(call
->a_args
.lock
.fh
));
329 call
->a_args
.lock
.caller
= system_utsname
.nodename
;
330 call
->a_args
.lock
.oh
.len
= lock
->oh
.len
;
332 /* set default data area */
333 call
->a_args
.lock
.oh
.data
= call
->a_owner
;
334 call
->a_args
.lock
.svid
= lock
->fl
.fl_pid
;
336 if (lock
->oh
.len
> NLMCLNT_OHSIZE
) {
337 void *data
= kmalloc(lock
->oh
.len
, GFP_KERNEL
);
340 call
->a_args
.lock
.oh
.data
= (u8
*) data
;
343 memcpy(call
->a_args
.lock
.oh
.data
, lock
->oh
.data
, lock
->oh
.len
);
347 static void nlmsvc_freegrantargs(struct nlm_rqst
*call
)
349 if (call
->a_args
.lock
.oh
.data
!= call
->a_owner
)
350 kfree(call
->a_args
.lock
.oh
.data
);
354 * Attempt to establish a lock, and if it can't be granted, block it
358 nlmsvc_lock(struct svc_rqst
*rqstp
, struct nlm_file
*file
,
359 struct nlm_lock
*lock
, int wait
, struct nlm_cookie
*cookie
)
361 struct nlm_block
*block
, *newblock
= NULL
;
365 dprintk("lockd: nlmsvc_lock(%s/%ld, ty=%d, pi=%d, %Ld-%Ld, bl=%d)\n",
366 file
->f_file
->f_dentry
->d_inode
->i_sb
->s_id
,
367 file
->f_file
->f_dentry
->d_inode
->i_ino
,
368 lock
->fl
.fl_type
, lock
->fl
.fl_pid
,
369 (long long)lock
->fl
.fl_start
,
370 (long long)lock
->fl
.fl_end
,
374 lock
->fl
.fl_flags
&= ~FL_SLEEP
;
376 /* Lock file against concurrent access */
378 /* Get existing block (in case client is busy-waiting) */
379 block
= nlmsvc_lookup_block(file
, lock
);
381 if (newblock
!= NULL
)
382 lock
= &newblock
->b_call
->a_args
.lock
;
384 lock
= &block
->b_call
->a_args
.lock
;
386 error
= posix_lock_file(file
->f_file
, &lock
->fl
);
387 lock
->fl
.fl_flags
&= ~FL_SLEEP
;
389 dprintk("lockd: posix_lock_file returned %d\n", error
);
400 default: /* includes ENOLCK */
401 ret
= nlm_lck_denied_nolocks
;
405 ret
= nlm_lck_denied
;
409 ret
= nlm_lck_blocked
;
413 /* If we don't have a block, create and initialize it. Then
414 * retry because we may have slept in kmalloc. */
415 /* We have to release f_sema as nlmsvc_create_block may try to
416 * to claim it while doing host garbage collection */
417 if (newblock
== NULL
) {
419 dprintk("lockd: blocking on this lock (allocating).\n");
420 if (!(newblock
= nlmsvc_create_block(rqstp
, file
, lock
, cookie
)))
421 return nlm_lck_denied_nolocks
;
425 /* Append to list of blocked */
426 nlmsvc_insert_block(newblock
, NLM_NEVER
);
429 nlmsvc_release_block(newblock
);
430 nlmsvc_release_block(block
);
431 dprintk("lockd: nlmsvc_lock returned %u\n", ret
);
436 * Test for presence of a conflicting lock.
439 nlmsvc_testlock(struct nlm_file
*file
, struct nlm_lock
*lock
,
440 struct nlm_lock
*conflock
)
442 dprintk("lockd: nlmsvc_testlock(%s/%ld, ty=%d, %Ld-%Ld)\n",
443 file
->f_file
->f_dentry
->d_inode
->i_sb
->s_id
,
444 file
->f_file
->f_dentry
->d_inode
->i_ino
,
446 (long long)lock
->fl
.fl_start
,
447 (long long)lock
->fl
.fl_end
);
449 if (posix_test_lock(file
->f_file
, &lock
->fl
, &conflock
->fl
)) {
450 dprintk("lockd: conflicting lock(ty=%d, %Ld-%Ld)\n",
451 conflock
->fl
.fl_type
,
452 (long long)conflock
->fl
.fl_start
,
453 (long long)conflock
->fl
.fl_end
);
454 conflock
->caller
= "somehost"; /* FIXME */
455 conflock
->oh
.len
= 0; /* don't return OH info */
456 conflock
->svid
= conflock
->fl
.fl_pid
;
457 return nlm_lck_denied
;
465 * This implies a CANCEL call: We send a GRANT_MSG, the client replies
466 * with a GRANT_RES call which gets lost, and calls UNLOCK immediately
467 * afterwards. In this case the block will still be there, and hence
471 nlmsvc_unlock(struct nlm_file
*file
, struct nlm_lock
*lock
)
475 dprintk("lockd: nlmsvc_unlock(%s/%ld, pi=%d, %Ld-%Ld)\n",
476 file
->f_file
->f_dentry
->d_inode
->i_sb
->s_id
,
477 file
->f_file
->f_dentry
->d_inode
->i_ino
,
479 (long long)lock
->fl
.fl_start
,
480 (long long)lock
->fl
.fl_end
);
482 /* First, cancel any lock that might be there */
483 nlmsvc_cancel_blocked(file
, lock
);
485 lock
->fl
.fl_type
= F_UNLCK
;
486 error
= posix_lock_file(file
->f_file
, &lock
->fl
);
488 return (error
< 0)? nlm_lck_denied_nolocks
: nlm_granted
;
492 * Cancel a previously blocked request.
494 * A cancel request always overrides any grant that may currently
496 * The calling procedure must check whether the file can be closed.
499 nlmsvc_cancel_blocked(struct nlm_file
*file
, struct nlm_lock
*lock
)
501 struct nlm_block
*block
;
504 dprintk("lockd: nlmsvc_cancel(%s/%ld, pi=%d, %Ld-%Ld)\n",
505 file
->f_file
->f_dentry
->d_inode
->i_sb
->s_id
,
506 file
->f_file
->f_dentry
->d_inode
->i_ino
,
508 (long long)lock
->fl
.fl_start
,
509 (long long)lock
->fl
.fl_end
);
512 block
= nlmsvc_lookup_block(file
, lock
);
515 status
= nlmsvc_unlink_block(block
);
516 nlmsvc_release_block(block
);
518 return status
? nlm_lck_denied
: nlm_granted
;
522 * Unblock a blocked lock request. This is a callback invoked from the
523 * VFS layer when a lock on which we blocked is removed.
525 * This function doesn't grant the blocked lock instantly, but rather moves
526 * the block to the head of nlm_blocked where it can be picked up by lockd.
529 nlmsvc_notify_blocked(struct file_lock
*fl
)
531 struct nlm_block
**bp
, *block
;
533 dprintk("lockd: VFS unblock notification for block %p\n", fl
);
534 for (bp
= &nlm_blocked
; (block
= *bp
) != 0; bp
= &block
->b_next
) {
535 if (nlm_compare_locks(&block
->b_call
->a_args
.lock
.fl
, fl
)) {
536 nlmsvc_insert_block(block
, 0);
537 svc_wake_up(block
->b_daemon
);
542 printk(KERN_WARNING
"lockd: notification for unknown block!\n");
545 static int nlmsvc_same_owner(struct file_lock
*fl1
, struct file_lock
*fl2
)
547 return fl1
->fl_owner
== fl2
->fl_owner
&& fl1
->fl_pid
== fl2
->fl_pid
;
550 struct lock_manager_operations nlmsvc_lock_operations
= {
551 .fl_compare_owner
= nlmsvc_same_owner
,
552 .fl_notify
= nlmsvc_notify_blocked
,
556 * Try to claim a lock that was previously blocked.
558 * Note that we use both the RPC_GRANTED_MSG call _and_ an async
559 * RPC thread when notifying the client. This seems like overkill...
561 * - we don't want to use a synchronous RPC thread, otherwise
562 * we might find ourselves hanging on a dead portmapper.
563 * - Some lockd implementations (e.g. HP) don't react to
564 * RPC_GRANTED calls; they seem to insist on RPC_GRANTED_MSG calls.
567 nlmsvc_grant_blocked(struct nlm_block
*block
)
569 struct nlm_file
*file
= block
->b_file
;
570 struct nlm_lock
*lock
= &block
->b_call
->a_args
.lock
;
573 dprintk("lockd: grant blocked lock %p\n", block
);
575 /* Unlink block request from list */
576 nlmsvc_unlink_block(block
);
578 /* If b_granted is true this means we've been here before.
579 * Just retry the grant callback, possibly refreshing the RPC
581 if (block
->b_granted
) {
582 nlm_rebind_host(block
->b_host
);
586 /* Try the lock operation again */
587 lock
->fl
.fl_flags
|= FL_SLEEP
;
588 error
= posix_lock_file(file
->f_file
, &lock
->fl
);
589 lock
->fl
.fl_flags
&= ~FL_SLEEP
;
595 dprintk("lockd: lock still blocked\n");
596 nlmsvc_insert_block(block
, NLM_NEVER
);
599 printk(KERN_WARNING
"lockd: unexpected error %d in %s!\n",
600 -error
, __FUNCTION__
);
601 nlmsvc_insert_block(block
, 10 * HZ
);
606 /* Lock was granted by VFS. */
607 dprintk("lockd: GRANTing blocked lock.\n");
608 block
->b_granted
= 1;
610 /* Schedule next grant callback in 30 seconds */
611 nlmsvc_insert_block(block
, 30 * HZ
);
613 /* Call the client */
614 kref_get(&block
->b_count
);
615 if (nlm_async_call(block
->b_call
, NLMPROC_GRANTED_MSG
,
616 &nlmsvc_grant_ops
) < 0)
617 nlmsvc_release_block(block
);
621 * This is the callback from the RPC layer when the NLM_GRANTED_MSG
622 * RPC call has succeeded or timed out.
623 * Like all RPC callbacks, it is invoked by the rpciod process, so it
624 * better not sleep. Therefore, we put the blocked lock on the nlm_blocked
625 * chain once more in order to have it removed by lockd itself (which can
626 * then sleep on the file semaphore without disrupting e.g. the nfs client).
628 static void nlmsvc_grant_callback(struct rpc_task
*task
, void *data
)
630 struct nlm_rqst
*call
= data
;
631 struct nlm_block
*block
= call
->a_block
;
632 unsigned long timeout
;
634 dprintk("lockd: GRANT_MSG RPC callback\n");
636 /* Technically, we should down the file semaphore here. Since we
637 * move the block towards the head of the queue only, no harm
638 * can be done, though. */
639 if (task
->tk_status
< 0) {
640 /* RPC error: Re-insert for retransmission */
642 } else if (block
->b_done
) {
643 /* Block already removed, kill it for real */
646 /* Call was successful, now wait for client callback */
649 nlmsvc_insert_block(block
, timeout
);
650 svc_wake_up(block
->b_daemon
);
653 static void nlmsvc_grant_release(void *data
)
655 struct nlm_rqst
*call
= data
;
657 nlmsvc_release_block(call
->a_block
);
660 static const struct rpc_call_ops nlmsvc_grant_ops
= {
661 .rpc_call_done
= nlmsvc_grant_callback
,
662 .rpc_release
= nlmsvc_grant_release
,
666 * We received a GRANT_RES callback. Try to find the corresponding
670 nlmsvc_grant_reply(struct svc_rqst
*rqstp
, struct nlm_cookie
*cookie
, u32 status
)
672 struct nlm_block
*block
;
673 struct nlm_file
*file
;
675 dprintk("grant_reply: looking for cookie %x, host (%08x), s=%d \n",
676 *(unsigned int *)(cookie
->data
),
677 ntohl(rqstp
->rq_addr
.sin_addr
.s_addr
), status
);
678 if (!(block
= nlmsvc_find_block(cookie
, &rqstp
->rq_addr
)))
680 file
= block
->b_file
;
683 if (status
== NLM_LCK_DENIED_GRACE_PERIOD
) {
684 /* Try again in a couple of seconds */
685 nlmsvc_insert_block(block
, 10 * HZ
);
687 /* Lock is now held by client, or has been rejected.
688 * In both cases, the block should be removed. */
689 nlmsvc_unlink_block(block
);
692 nlmsvc_release_block(block
);
696 * Retry all blocked locks that have been notified. This is where lockd
697 * picks up locks that can be granted, or grant notifications that must
701 nlmsvc_retry_blocked(void)
703 struct nlm_block
*block
;
705 dprintk("nlmsvc_retry_blocked(%p, when=%ld)\n",
707 nlm_blocked
? nlm_blocked
->b_when
: 0);
708 while ((block
= nlm_blocked
) != 0) {
709 if (block
->b_when
== NLM_NEVER
)
711 if (time_after(block
->b_when
,jiffies
))
713 dprintk("nlmsvc_retry_blocked(%p, when=%ld, done=%d)\n",
714 block
, block
->b_when
, block
->b_done
);
715 kref_get(&block
->b_count
);
717 nlmsvc_unlink_block(block
);
719 nlmsvc_grant_blocked(block
);
720 nlmsvc_release_block(block
);
723 if ((block
= nlm_blocked
) && block
->b_when
!= NLM_NEVER
)
724 return (block
->b_when
- jiffies
);
726 return MAX_SCHEDULE_TIMEOUT
;