4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2003 Hewlett-Packard Development Company LP.
24 * Developed under the sponsorship of the US Government under
25 * Subcontract No. B514193
27 * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
30 * Copyright (c) 2010, 2017, Intel Corporation.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
38 * This file implements POSIX lock type for Lustre.
39 * Its policy properties are start and end of extent and PID.
41 * These locks are only done through MDS due to POSIX semantics requiring
42 * e.g. that locks could be only partially released and as such split into
43 * two parts, and also that two adjacent locks from the same process may be
44 * merged into a single wider lock.
46 * Lock modes are mapped like this:
47 * PR and PW for READ and WRITE locks
48 * NL to request a releasing of a portion of the lock
50 * These flock locks never timeout.
53 #define DEBUG_SUBSYSTEM S_LDLM
55 #include <linux/list.h>
56 #include <lustre_dlm.h>
57 #include <obd_support.h>
58 #include <obd_class.h>
59 #include <lustre_lib.h>
61 #include "ldlm_internal.h"
63 int ldlm_flock_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
64 void *data, int flag);
67 * list_for_remaining_safe - iterate over the remaining entries in a list
68 * and safeguard against removal of a list entry.
69 * \param pos the &struct list_head to use as a loop counter. pos MUST
70 * have been initialized prior to using it in this macro.
71 * \param n another &struct list_head to use as temporary storage
72 * \param head the head for your list.
74 #define list_for_remaining_safe(pos, n, head) \
75 for (n = pos->next; pos != (head); pos = n, n = pos->next)
78 ldlm_same_flock_owner(struct ldlm_lock *lock, struct ldlm_lock *new)
80 return ((new->l_policy_data.l_flock.owner ==
81 lock->l_policy_data.l_flock.owner) &&
82 (new->l_export == lock->l_export));
86 ldlm_flocks_overlap(struct ldlm_lock *lock, struct ldlm_lock *new)
88 return ((new->l_policy_data.l_flock.start <=
89 lock->l_policy_data.l_flock.end) &&
90 (new->l_policy_data.l_flock.end >=
91 lock->l_policy_data.l_flock.start));
94 static inline void ldlm_flock_blocking_link(struct ldlm_lock *req,
95 struct ldlm_lock *lock)
98 if (req->l_export == NULL)
101 LASSERT(hlist_unhashed(&req->l_exp_flock_hash));
103 req->l_policy_data.l_flock.blocking_owner =
104 lock->l_policy_data.l_flock.owner;
105 req->l_policy_data.l_flock.blocking_export =
107 atomic_set(&req->l_policy_data.l_flock.blocking_refs, 0);
109 cfs_hash_add(req->l_export->exp_flock_hash,
110 &req->l_policy_data.l_flock.owner,
111 &req->l_exp_flock_hash);
114 static inline void ldlm_flock_blocking_unlink(struct ldlm_lock *req)
116 /* For server only */
117 if (req->l_export == NULL)
120 check_res_locked(req->l_resource);
121 if (req->l_export->exp_flock_hash != NULL &&
122 !hlist_unhashed(&req->l_exp_flock_hash))
123 cfs_hash_del(req->l_export->exp_flock_hash,
124 &req->l_policy_data.l_flock.owner,
125 &req->l_exp_flock_hash);
129 ldlm_flock_destroy(struct ldlm_lock *lock, enum ldlm_mode mode, __u64 flags)
133 LDLM_DEBUG(lock, "ldlm_flock_destroy(mode: %d, flags: %#llx)",
136 /* Safe to not lock here, since it should be empty anyway */
137 LASSERT(hlist_unhashed(&lock->l_exp_flock_hash));
139 list_del_init(&lock->l_res_link);
140 if (flags == LDLM_FL_WAIT_NOREPROC) {
141 /* client side - set a flag to prevent sending a CANCEL */
142 lock->l_flags |= LDLM_FL_LOCAL_ONLY | LDLM_FL_CBPENDING;
144 /* when reaching here, it is under lock_res_and_lock(). Thus,
145 * need call the nolock version of ldlm_lock_decref_internal
147 ldlm_lock_decref_internal_nolock(lock, mode);
150 ldlm_lock_destroy_nolock(lock);
155 * POSIX locks deadlock detection code.
157 * Given a new lock \a req and an existing lock \a bl_lock it conflicts
158 * with, we need to iterate through all blocked POSIX locks for this
159 * export and see if there is a deadlock condition arising. (i.e. when
160 * one client holds a lock on something and want a lock on something
161 * else and at the same time another client has the opposite situation).
164 struct ldlm_flock_lookup_cb_data {
166 struct ldlm_lock *lock;
167 struct obd_export *exp;
170 static int ldlm_flock_lookup_cb(struct cfs_hash *hs, struct cfs_hash_bd *bd,
171 struct hlist_node *hnode, void *data)
173 struct ldlm_flock_lookup_cb_data *cb_data = data;
174 struct obd_export *exp = cfs_hash_object(hs, hnode);
175 struct ldlm_lock *lock;
177 lock = cfs_hash_lookup(exp->exp_flock_hash, cb_data->bl_owner);
181 /* Stop on first found lock. Same process can't sleep twice */
182 cb_data->lock = lock;
183 cb_data->exp = class_export_get(exp);
189 ldlm_flock_deadlock(struct ldlm_lock *req, struct ldlm_lock *bl_lock)
191 struct obd_export *req_exp = req->l_export;
192 struct obd_export *bl_exp = bl_lock->l_export;
193 __u64 req_owner = req->l_policy_data.l_flock.owner;
194 __u64 bl_owner = bl_lock->l_policy_data.l_flock.owner;
196 /* For server only */
200 class_export_get(bl_exp);
202 struct ldlm_flock_lookup_cb_data cb_data = {
203 .bl_owner = &bl_owner,
207 struct obd_export *bl_exp_new;
208 struct ldlm_lock *lock = NULL;
209 struct ldlm_flock *flock;
211 if (bl_exp->exp_flock_hash != NULL) {
212 cfs_hash_for_each_key(
213 bl_exp->exp_obd->obd_nid_hash,
214 &bl_exp->exp_connection->c_peer.nid,
215 ldlm_flock_lookup_cb, &cb_data);
221 class_export_put(bl_exp);
222 bl_exp = cb_data.exp;
224 LASSERT(req != lock);
225 flock = &lock->l_policy_data.l_flock;
226 LASSERT(flock->owner == bl_owner);
227 bl_owner = flock->blocking_owner;
228 bl_exp_new = class_export_get(flock->blocking_export);
229 class_export_put(bl_exp);
231 cfs_hash_put(bl_exp->exp_flock_hash, &lock->l_exp_flock_hash);
234 if (bl_exp->exp_failed)
237 if (bl_owner == req_owner &&
238 (bl_exp->exp_connection->c_peer.nid ==
239 req_exp->exp_connection->c_peer.nid)) {
240 class_export_put(bl_exp);
244 class_export_put(bl_exp);
249 static void ldlm_flock_cancel_on_deadlock(struct ldlm_lock *lock,
250 struct list_head *work_list)
252 CDEBUG(D_INFO, "reprocess deadlock req=%p\n", lock);
254 if ((exp_connect_flags(lock->l_export) &
255 OBD_CONNECT_FLOCK_DEAD) == 0) {
256 CERROR("deadlock found, but client doesn't support flock canceliation\n");
258 LASSERT(lock->l_completion_ast);
259 LASSERT(!ldlm_is_ast_sent(lock));
260 lock->l_flags |= (LDLM_FL_AST_SENT | LDLM_FL_CANCEL_ON_BLOCK |
261 LDLM_FL_FLOCK_DEADLOCK);
262 ldlm_flock_blocking_unlink(lock);
263 ldlm_resource_unlink_lock(lock);
264 ldlm_add_ast_work_item(lock, NULL, work_list);
269 * Process a granting attempt for flock lock.
270 * Must be called under ns lock held.
272 * This function looks for any conflicts for \a lock in the granted or
273 * waiting queues. The lock is granted if no conflicts are found in
277 ldlm_process_flock_lock(struct ldlm_lock *req, __u64 *flags,
278 enum ldlm_process_intention intention,
279 enum ldlm_error *err, struct list_head *work_list)
281 struct ldlm_resource *res = req->l_resource;
282 struct ldlm_namespace *ns = ldlm_res_to_ns(res);
283 struct list_head *tmp;
284 struct list_head *ownlocks = NULL;
285 struct ldlm_lock *lock = NULL;
286 struct ldlm_lock *new = req;
287 struct ldlm_lock *new2 = NULL;
288 enum ldlm_mode mode = req->l_req_mode;
289 int local = ns_is_client(ns);
290 int added = (mode == LCK_NL);
293 const struct ldlm_callback_suite null_cbs = { NULL };
294 struct list_head *grant_work = (intention == LDLM_PROCESS_ENQUEUE ?
298 CDEBUG(D_DLMTRACE, "flags %#llx owner %llu pid %u mode %u start "
299 "%llu end %llu\n", *flags,
300 new->l_policy_data.l_flock.owner,
301 new->l_policy_data.l_flock.pid, mode,
302 req->l_policy_data.l_flock.start,
303 req->l_policy_data.l_flock.end);
308 /* No blocking ASTs are sent to the clients for
309 * Posix file & record locks
311 req->l_blocking_ast = NULL;
313 /* Called on the server for lock cancels. */
314 req->l_blocking_ast = ldlm_flock_blocking_ast;
318 if ((*flags == LDLM_FL_WAIT_NOREPROC) || (mode == LCK_NL)) {
319 /* This loop determines where this processes locks start
320 * in the resource lr_granted list.
322 list_for_each(tmp, &res->lr_granted) {
323 lock = list_entry(tmp, struct ldlm_lock,
325 if (ldlm_same_flock_owner(lock, req)) {
331 int reprocess_failed = 0;
332 lockmode_verify(mode);
334 /* This loop determines if there are existing locks
335 * that conflict with the new lock request.
337 list_for_each(tmp, &res->lr_granted) {
338 lock = list_entry(tmp, struct ldlm_lock,
341 if (ldlm_same_flock_owner(lock, req)) {
347 /* locks are compatible, overlap doesn't matter */
348 if (lockmode_compat(lock->l_granted_mode, mode))
351 if (!ldlm_flocks_overlap(lock, req))
354 if (intention != LDLM_PROCESS_ENQUEUE) {
355 reprocess_failed = 1;
356 if (ldlm_flock_deadlock(req, lock)) {
357 ldlm_flock_cancel_on_deadlock(
359 RETURN(LDLM_ITER_CONTINUE);
364 if (*flags & LDLM_FL_BLOCK_NOWAIT) {
365 ldlm_flock_destroy(req, mode, *flags);
367 RETURN(LDLM_ITER_STOP);
370 if (*flags & LDLM_FL_TEST_LOCK) {
371 ldlm_flock_destroy(req, mode, *flags);
372 req->l_req_mode = lock->l_granted_mode;
373 req->l_policy_data.l_flock.pid =
374 lock->l_policy_data.l_flock.pid;
375 req->l_policy_data.l_flock.start =
376 lock->l_policy_data.l_flock.start;
377 req->l_policy_data.l_flock.end =
378 lock->l_policy_data.l_flock.end;
379 *flags |= LDLM_FL_LOCK_CHANGED;
380 RETURN(LDLM_ITER_STOP);
383 /* add lock to blocking list before deadlock
384 * check to prevent race
386 ldlm_flock_blocking_link(req, lock);
388 if (ldlm_flock_deadlock(req, lock)) {
389 ldlm_flock_blocking_unlink(req);
390 ldlm_flock_destroy(req, mode, *flags);
392 RETURN(LDLM_ITER_STOP);
395 ldlm_resource_add_lock(res, &res->lr_waiting, req);
396 *flags |= LDLM_FL_BLOCK_GRANTED;
397 RETURN(LDLM_ITER_STOP);
399 if (reprocess_failed)
400 RETURN(LDLM_ITER_CONTINUE);
403 if (*flags & LDLM_FL_TEST_LOCK) {
404 ldlm_flock_destroy(req, mode, *flags);
405 req->l_req_mode = LCK_NL;
406 *flags |= LDLM_FL_LOCK_CHANGED;
407 RETURN(LDLM_ITER_STOP);
410 /* In case we had slept on this lock request take it off of the
411 * deadlock detection hash list.
413 ldlm_flock_blocking_unlink(req);
415 /* Scan the locks owned by this process that overlap this request.
416 * We may have to merge or split existing locks.
420 ownlocks = &res->lr_granted;
422 list_for_remaining_safe(ownlocks, tmp, &res->lr_granted) {
423 lock = list_entry(ownlocks, struct ldlm_lock, l_res_link);
425 if (!ldlm_same_flock_owner(lock, new))
428 if (lock->l_granted_mode == mode) {
429 /* If the modes are the same then we need to process
430 * locks that overlap OR adjoin the new lock. The extra
431 * logic condition is necessary to deal with arithmetic
432 * overflow and underflow.
434 if ((new->l_policy_data.l_flock.start >
435 (lock->l_policy_data.l_flock.end + 1))
436 && (lock->l_policy_data.l_flock.end !=
440 if ((new->l_policy_data.l_flock.end <
441 (lock->l_policy_data.l_flock.start - 1))
442 && (lock->l_policy_data.l_flock.start != 0))
445 if (new->l_policy_data.l_flock.start <
446 lock->l_policy_data.l_flock.start) {
447 lock->l_policy_data.l_flock.start =
448 new->l_policy_data.l_flock.start;
450 new->l_policy_data.l_flock.start =
451 lock->l_policy_data.l_flock.start;
454 if (new->l_policy_data.l_flock.end >
455 lock->l_policy_data.l_flock.end) {
456 lock->l_policy_data.l_flock.end =
457 new->l_policy_data.l_flock.end;
459 new->l_policy_data.l_flock.end =
460 lock->l_policy_data.l_flock.end;
464 ldlm_flock_destroy(lock, mode, *flags);
472 if (new->l_policy_data.l_flock.start >
473 lock->l_policy_data.l_flock.end)
476 if (new->l_policy_data.l_flock.end <
477 lock->l_policy_data.l_flock.start)
482 if (new->l_policy_data.l_flock.start <=
483 lock->l_policy_data.l_flock.start) {
484 if (new->l_policy_data.l_flock.end <
485 lock->l_policy_data.l_flock.end) {
486 lock->l_policy_data.l_flock.start =
487 new->l_policy_data.l_flock.end + 1;
490 ldlm_flock_destroy(lock, lock->l_req_mode, *flags);
493 if (new->l_policy_data.l_flock.end >=
494 lock->l_policy_data.l_flock.end) {
495 lock->l_policy_data.l_flock.end =
496 new->l_policy_data.l_flock.start - 1;
500 /* split the existing lock into two locks */
502 /* if this is an F_UNLCK operation then we could avoid
503 * allocating a new lock and use the req lock passed in
504 * with the request but this would complicate the reply
505 * processing since updates to req get reflected in the
506 * reply. The client side replays the lock request so
507 * it must see the original lock data in the reply.
510 /* XXX - if ldlm_lock_new() can sleep we should
511 * release the lr_lock, allocate the new lock,
512 * and restart processing this lock.
515 unlock_res_and_lock(req);
516 new2 = ldlm_lock_create(ns, &res->lr_name, LDLM_FLOCK,
517 lock->l_granted_mode, &null_cbs,
518 NULL, 0, LVB_T_NONE);
519 lock_res_and_lock(req);
521 ldlm_flock_destroy(req, lock->l_granted_mode,
523 *err = PTR_ERR(new2);
524 RETURN(LDLM_ITER_STOP);
531 new2->l_granted_mode = lock->l_granted_mode;
532 new2->l_policy_data.l_flock.pid =
533 new->l_policy_data.l_flock.pid;
534 new2->l_policy_data.l_flock.owner =
535 new->l_policy_data.l_flock.owner;
536 new2->l_policy_data.l_flock.start =
537 lock->l_policy_data.l_flock.start;
538 new2->l_policy_data.l_flock.end =
539 new->l_policy_data.l_flock.start - 1;
540 lock->l_policy_data.l_flock.start =
541 new->l_policy_data.l_flock.end + 1;
542 new2->l_conn_export = lock->l_conn_export;
543 if (lock->l_export != NULL) {
544 new2->l_export = class_export_lock_get(lock->l_export,
546 if (new2->l_export->exp_lock_hash &&
547 hlist_unhashed(&new2->l_exp_hash))
548 cfs_hash_add(new2->l_export->exp_lock_hash,
549 &new2->l_remote_handle,
552 if (*flags == LDLM_FL_WAIT_NOREPROC)
553 ldlm_lock_addref_internal_nolock(new2,
554 lock->l_granted_mode);
556 /* insert new2 at lock */
557 ldlm_resource_add_lock(res, ownlocks, new2);
558 LDLM_LOCK_RELEASE(new2);
562 /* if new2 is created but never used, destroy it*/
563 if (splitted == 0 && new2 != NULL)
564 ldlm_lock_destroy_nolock(new2);
566 /* At this point we're granting the lock request. */
567 req->l_granted_mode = req->l_req_mode;
569 /* Add req to the granted queue before calling ldlm_reprocess_all(). */
571 list_del_init(&req->l_res_link);
572 /* insert new lock before ownlocks in list. */
573 ldlm_resource_add_lock(res, ownlocks, req);
576 if (*flags != LDLM_FL_WAIT_NOREPROC) {
577 #ifdef HAVE_SERVER_SUPPORT
578 if (intention == LDLM_PROCESS_ENQUEUE) {
579 /* If this is an unlock, reprocess the waitq and
580 * send completions ASTs for locks that can now be
581 * granted. The only problem with doing this
582 * reprocessing here is that the completion ASTs for
583 * newly granted locks will be sent before the unlock
584 * completion is sent. It shouldn't be an issue. Also
585 * note that ldlm_process_flock_lock() will recurse,
586 * but only once because 'intention' won't be
587 * LDLM_PROCESS_ENQUEUE from ldlm_reprocess_queue.
589 if ((mode == LCK_NL) && overlaps) {
590 struct list_head rpc_list;
593 INIT_LIST_HEAD(&rpc_list);
595 ldlm_reprocess_queue(res, &res->lr_waiting,
597 LDLM_PROCESS_RESCAN, NULL);
599 unlock_res_and_lock(req);
600 rc = ldlm_run_ast_work(ns, &rpc_list,
602 lock_res_and_lock(req);
607 LASSERT(req->l_completion_ast);
608 ldlm_add_ast_work_item(req, NULL, grant_work);
610 #else /* !HAVE_SERVER_SUPPORT */
611 /* The only one possible case for client-side calls flock
612 * policy function is ldlm_flock_completion_ast inside which
613 * carries LDLM_FL_WAIT_NOREPROC flag.
615 CERROR("Illegal parameter for client-side-only module.\n");
617 #endif /* HAVE_SERVER_SUPPORT */
620 /* In case we're reprocessing the requested lock we can't destroy
621 * it until after calling ldlm_add_ast_work_item() above so that laawi()
622 * can bump the reference count on \a req. Otherwise \a req
623 * could be freed before the completion AST can be sent.
626 ldlm_flock_destroy(req, mode, *flags);
628 ldlm_resource_dump(D_INFO, res);
629 RETURN(LDLM_ITER_CONTINUE);
632 struct ldlm_flock_wait_data {
633 struct ldlm_lock *fwd_lock;
638 ldlm_flock_interrupted_wait(void *data)
640 struct ldlm_lock *lock;
643 lock = ((struct ldlm_flock_wait_data *)data)->fwd_lock;
645 /* take lock off the deadlock detection hash list.
647 lock_res_and_lock(lock);
648 ldlm_flock_blocking_unlink(lock);
650 /* client side - set flag to prevent lock from being put on LRU list */
651 ldlm_set_cbpending(lock);
652 unlock_res_and_lock(lock);
658 * Flock completion callback function.
660 * \param lock [in,out]: A lock to be handled
661 * \param flags [in]: flags
662 * \param *data [in]: ldlm_work_cp_ast_lock() will use ldlm_cb_set_arg
664 * \retval 0 : success
665 * \retval <0 : failure
668 ldlm_flock_completion_ast(struct ldlm_lock *lock, __u64 flags, void *data)
670 struct file_lock *getlk = lock->l_ast_data;
671 struct obd_device *obd;
672 struct obd_import *imp = NULL;
673 struct ldlm_flock_wait_data fwd;
674 struct l_wait_info lwi;
679 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_CP_CB_WAIT2, 4);
680 if (OBD_FAIL_PRECHECK(OBD_FAIL_LDLM_CP_CB_WAIT3)) {
681 lock_res_and_lock(lock);
682 lock->l_flags |= LDLM_FL_FAIL_LOC;
683 unlock_res_and_lock(lock);
684 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_CP_CB_WAIT3, 4);
686 CDEBUG(D_DLMTRACE, "flags: %#llx data: %p getlk: %p\n",
689 LASSERT(flags != LDLM_FL_WAIT_NOREPROC);
691 if (flags & LDLM_FL_FAILED)
694 if (!(flags & LDLM_FL_BLOCKED_MASK)) {
696 /* mds granted the lock in the reply */
698 /* CP AST RPC: lock get granted, wake it up */
699 wake_up(&lock->l_waitq);
704 "client-side enqueue returned a blocked lock, sleeping");
706 obd = class_exp2obd(lock->l_conn_export);
708 /* if this is a local lock, there is no import */
710 imp = obd->u.cli.cl_import;
713 spin_lock(&imp->imp_lock);
714 fwd.fwd_generation = imp->imp_generation;
715 spin_unlock(&imp->imp_lock);
718 lwi = LWI_TIMEOUT_INTR(0, NULL, ldlm_flock_interrupted_wait, &fwd);
720 /* Go to sleep until the lock is granted. */
721 rc = l_wait_event(lock->l_waitq, is_granted_or_cancelled(lock), &lwi);
724 LDLM_DEBUG(lock, "client-side enqueue waking up: failed (%d)",
730 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_CP_CB_WAIT, 10);
732 if (OBD_FAIL_PRECHECK(OBD_FAIL_LDLM_CP_CB_WAIT4)) {
733 lock_res_and_lock(lock);
734 /* DEADLOCK is always set with CBPENDING */
735 lock->l_flags |= LDLM_FL_FLOCK_DEADLOCK | LDLM_FL_CBPENDING;
736 unlock_res_and_lock(lock);
737 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_CP_CB_WAIT4, 4);
739 if (OBD_FAIL_PRECHECK(OBD_FAIL_LDLM_CP_CB_WAIT5)) {
740 lock_res_and_lock(lock);
741 /* DEADLOCK is always set with CBPENDING */
742 lock->l_flags |= (LDLM_FL_FAIL_LOC |
743 LDLM_FL_FLOCK_DEADLOCK | LDLM_FL_CBPENDING);
744 unlock_res_and_lock(lock);
745 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_CP_CB_WAIT5, 4);
748 lock_res_and_lock(lock);
751 /* Protect against race where lock could have been just destroyed
752 * due to overlap in ldlm_process_flock_lock().
754 if (ldlm_is_destroyed(lock)) {
755 unlock_res_and_lock(lock);
756 LDLM_DEBUG(lock, "client-side enqueue waking up: destroyed");
758 /* An error is still to be returned, to propagate it up to
759 * ldlm_cli_enqueue_fini() caller. */
763 /* ldlm_lock_enqueue() has already placed lock on the granted list. */
764 ldlm_resource_unlink_lock(lock);
766 /* Import invalidation. We need to actually release the lock
767 * references being held, so that it can go away. No point in
768 * holding the lock even if app still believes it has it, since
769 * server already dropped it anyway. Only for granted locks too.
771 /* Do the same for DEADLOCK'ed locks. */
772 if (ldlm_is_failed(lock) || ldlm_is_flock_deadlock(lock)) {
775 if (flags & LDLM_FL_TEST_LOCK)
776 LASSERT(ldlm_is_test_lock(lock));
778 if (ldlm_is_test_lock(lock) || ldlm_is_flock_deadlock(lock))
779 mode = getlk->fl_type;
781 mode = lock->l_granted_mode;
783 if (ldlm_is_flock_deadlock(lock)) {
784 LDLM_DEBUG(lock, "client-side enqueue deadlock "
788 ldlm_flock_destroy(lock, mode, LDLM_FL_WAIT_NOREPROC);
789 unlock_res_and_lock(lock);
791 /* Need to wake up the waiter if we were evicted */
792 wake_up(&lock->l_waitq);
794 /* An error is still to be returned, to propagate it up to
795 * ldlm_cli_enqueue_fini() caller.
800 LDLM_DEBUG(lock, "client-side enqueue granted");
802 if (flags & LDLM_FL_TEST_LOCK) {
804 * fcntl(F_GETLK) request
805 * The old mode was saved in getlk->fl_type so that if the mode
806 * in the lock changes we can decref the appropriate refcount.
808 LASSERT(ldlm_is_test_lock(lock));
809 ldlm_flock_destroy(lock, getlk->fl_type, LDLM_FL_WAIT_NOREPROC);
810 switch (lock->l_granted_mode) {
812 getlk->fl_type = F_RDLCK;
815 getlk->fl_type = F_WRLCK;
818 getlk->fl_type = F_UNLCK;
820 getlk->fl_pid = (pid_t)lock->l_policy_data.l_flock.pid;
821 getlk->fl_start = (loff_t)lock->l_policy_data.l_flock.start;
822 getlk->fl_end = (loff_t)lock->l_policy_data.l_flock.end;
824 __u64 noreproc = LDLM_FL_WAIT_NOREPROC;
826 /* We need to reprocess the lock to do merges or splits
827 * with existing locks owned by this process.
829 ldlm_process_flock_lock(lock, &noreproc, 1, &err, NULL);
831 unlock_res_and_lock(lock);
834 EXPORT_SYMBOL(ldlm_flock_completion_ast);
836 int ldlm_flock_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
837 void *data, int flag)
842 LASSERT(flag == LDLM_CB_CANCELING);
844 /* take lock off the deadlock detection hash list. */
845 lock_res_and_lock(lock);
846 ldlm_flock_blocking_unlink(lock);
847 unlock_res_and_lock(lock);
851 void ldlm_flock_policy_wire_to_local(const union ldlm_wire_policy_data *wpolicy,
852 union ldlm_policy_data *lpolicy)
854 lpolicy->l_flock.start = wpolicy->l_flock.lfw_start;
855 lpolicy->l_flock.end = wpolicy->l_flock.lfw_end;
856 lpolicy->l_flock.pid = wpolicy->l_flock.lfw_pid;
857 lpolicy->l_flock.owner = wpolicy->l_flock.lfw_owner;
860 void ldlm_flock_policy_local_to_wire(const union ldlm_policy_data *lpolicy,
861 union ldlm_wire_policy_data *wpolicy)
863 memset(wpolicy, 0, sizeof(*wpolicy));
864 wpolicy->l_flock.lfw_start = lpolicy->l_flock.start;
865 wpolicy->l_flock.lfw_end = lpolicy->l_flock.end;
866 wpolicy->l_flock.lfw_pid = lpolicy->l_flock.pid;
867 wpolicy->l_flock.lfw_owner = lpolicy->l_flock.owner;
871 * Export handle<->flock hash operations.
874 ldlm_export_flock_hash(struct cfs_hash *hs, const void *key, unsigned mask)
876 return cfs_hash_u64_hash(*(__u64 *)key, mask);
880 ldlm_export_flock_key(struct hlist_node *hnode)
882 struct ldlm_lock *lock;
884 lock = hlist_entry(hnode, struct ldlm_lock, l_exp_flock_hash);
885 return &lock->l_policy_data.l_flock.owner;
889 ldlm_export_flock_keycmp(const void *key, struct hlist_node *hnode)
891 return !memcmp(ldlm_export_flock_key(hnode), key, sizeof(__u64));
895 ldlm_export_flock_object(struct hlist_node *hnode)
897 return hlist_entry(hnode, struct ldlm_lock, l_exp_flock_hash);
901 ldlm_export_flock_get(struct cfs_hash *hs, struct hlist_node *hnode)
903 struct ldlm_lock *lock;
904 struct ldlm_flock *flock;
906 lock = hlist_entry(hnode, struct ldlm_lock, l_exp_flock_hash);
909 flock = &lock->l_policy_data.l_flock;
910 LASSERT(flock->blocking_export != NULL);
911 class_export_get(flock->blocking_export);
912 atomic_inc(&flock->blocking_refs);
916 ldlm_export_flock_put(struct cfs_hash *hs, struct hlist_node *hnode)
918 struct ldlm_lock *lock;
919 struct ldlm_flock *flock;
921 lock = hlist_entry(hnode, struct ldlm_lock, l_exp_flock_hash);
923 flock = &lock->l_policy_data.l_flock;
924 LASSERT(flock->blocking_export != NULL);
925 class_export_put(flock->blocking_export);
926 if (atomic_dec_and_test(&flock->blocking_refs)) {
927 flock->blocking_owner = 0;
928 flock->blocking_export = NULL;
930 LDLM_LOCK_RELEASE(lock);
933 static struct cfs_hash_ops ldlm_export_flock_ops = {
934 .hs_hash = ldlm_export_flock_hash,
935 .hs_key = ldlm_export_flock_key,
936 .hs_keycmp = ldlm_export_flock_keycmp,
937 .hs_object = ldlm_export_flock_object,
938 .hs_get = ldlm_export_flock_get,
939 .hs_put = ldlm_export_flock_put,
940 .hs_put_locked = ldlm_export_flock_put,
943 int ldlm_init_flock_export(struct obd_export *exp)
945 if( strcmp(exp->exp_obd->obd_type->typ_name, LUSTRE_MDT_NAME) != 0)
948 exp->exp_flock_hash =
949 cfs_hash_create(obd_uuid2str(&exp->exp_client_uuid),
950 HASH_EXP_LOCK_CUR_BITS,
951 HASH_EXP_LOCK_MAX_BITS,
952 HASH_EXP_LOCK_BKT_BITS, 0,
953 CFS_HASH_MIN_THETA, CFS_HASH_MAX_THETA,
954 &ldlm_export_flock_ops,
955 CFS_HASH_DEFAULT | CFS_HASH_NBLK_CHANGE);
956 if (!exp->exp_flock_hash)
962 void ldlm_destroy_flock_export(struct obd_export *exp)
965 if (exp->exp_flock_hash) {
966 cfs_hash_putref(exp->exp_flock_hash);
967 exp->exp_flock_hash = NULL;