1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
4 * Copyright (c) 2002, 2003 Cluster File Systems, Inc.
5 * Author: Peter Braam <braam@clusterfs.com>
6 * Author: Phil Schwan <phil@clusterfs.com>
8 * This file is part of the Lustre file system, http://www.lustre.org
9 * Lustre is a trademark of Cluster File Systems, Inc.
11 * You may have signed or agreed to another license before downloading
12 * this software. If so, you are bound by the terms and conditions
13 * of that agreement, and the following does not apply to you. See the
14 * LICENSE file included with this distribution for more information.
16 * If you did not agree to a different license, then this copy of Lustre
17 * is open source software; you can redistribute it and/or modify it
18 * under the terms of version 2 of the GNU General Public License as
19 * published by the Free Software Foundation.
21 * In either case, Lustre is distributed in the hope that it will be
22 * useful, but WITHOUT ANY WARRANTY; without even the implied warranty
23 * of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
24 * license text for more details.
27 #define DEBUG_SUBSYSTEM S_LDLM
30 # include <libcfs/libcfs.h>
31 # ifndef HAVE_VFS_INTENT_PATCHES
32 # include <linux/lustre_intent.h>
35 # include <liblustre.h>
36 # include <libcfs/kp30.h>
39 #include <obd_class.h>
40 #include "ldlm_internal.h"
43 char *ldlm_lockname[] = {
54 char *ldlm_typename[] = {
61 char *ldlm_it2str(int it)
68 case (IT_OPEN | IT_CREAT):
81 CERROR("Unknown intent %d\n", it);
86 extern cfs_mem_cache_t *ldlm_lock_slab;
88 static ldlm_processing_policy ldlm_processing_policy_table[] = {
89 [LDLM_PLAIN] ldlm_process_plain_lock,
90 [LDLM_EXTENT] ldlm_process_extent_lock,
92 [LDLM_FLOCK] ldlm_process_flock_lock,
94 [LDLM_IBITS] ldlm_process_inodebits_lock,
97 ldlm_processing_policy ldlm_get_processing_policy(struct ldlm_resource *res)
99 return ldlm_processing_policy_table[res->lr_type];
102 void ldlm_register_intent(struct ldlm_namespace *ns, ldlm_res_policy arg)
108 * REFCOUNTED LOCK OBJECTS
113 * Lock refcounts, during creation:
114 * - one special one for allocation, dec'd only once in destroy
115 * - one for being a lock that's in-use
116 * - one for the addref associated with a new lock
118 struct ldlm_lock *ldlm_lock_get(struct ldlm_lock *lock)
120 atomic_inc(&lock->l_refc);
124 static void ldlm_lock_free(struct ldlm_lock *lock, size_t size)
126 LASSERT(size == sizeof(*lock));
127 OBD_SLAB_FREE(lock, ldlm_lock_slab, sizeof(*lock));
130 void ldlm_lock_put(struct ldlm_lock *lock)
134 LASSERT(lock->l_resource != LP_POISON);
135 LASSERT(atomic_read(&lock->l_refc) > 0);
136 if (atomic_dec_and_test(&lock->l_refc)) {
137 struct ldlm_resource *res;
140 "final lock_put on destroyed lock, freeing it.");
142 res = lock->l_resource;
143 LASSERT(lock->l_destroyed);
144 LASSERT(list_empty(&lock->l_res_link));
145 LASSERT(list_empty(&lock->l_pending_chain));
147 atomic_dec(&res->lr_namespace->ns_locks);
148 ldlm_resource_putref(res);
149 lock->l_resource = NULL;
150 if (lock->l_export) {
151 class_export_put(lock->l_export);
152 lock->l_export = NULL;
155 if (lock->l_lvb_data != NULL)
156 OBD_FREE(lock->l_lvb_data, lock->l_lvb_len);
158 ldlm_interval_free(ldlm_interval_detach(lock));
159 OBD_FREE_RCU_CB(lock, sizeof(*lock), &lock->l_handle,
166 int ldlm_lock_remove_from_lru_nolock(struct ldlm_lock *lock)
169 if (!list_empty(&lock->l_lru)) {
170 struct ldlm_namespace *ns = lock->l_resource->lr_namespace;
171 LASSERT(lock->l_resource->lr_type != LDLM_FLOCK);
172 list_del_init(&lock->l_lru);
174 LASSERT(ns->ns_nr_unused >= 0);
180 int ldlm_lock_remove_from_lru(struct ldlm_lock *lock)
182 struct ldlm_namespace *ns = lock->l_resource->lr_namespace;
185 spin_lock(&ns->ns_unused_lock);
186 rc = ldlm_lock_remove_from_lru_nolock(lock);
187 spin_unlock(&ns->ns_unused_lock);
192 void ldlm_lock_add_to_lru_nolock(struct ldlm_lock *lock)
194 struct ldlm_namespace *ns = lock->l_resource->lr_namespace;
195 lock->l_last_used = cfs_time_current();
196 LASSERT(list_empty(&lock->l_lru));
197 list_add_tail(&lock->l_lru, &ns->ns_unused_list);
198 LASSERT(ns->ns_nr_unused >= 0);
202 void ldlm_lock_add_to_lru(struct ldlm_lock *lock)
204 struct ldlm_namespace *ns = lock->l_resource->lr_namespace;
206 spin_lock(&ns->ns_unused_lock);
207 ldlm_lock_add_to_lru_nolock(lock);
208 spin_unlock(&ns->ns_unused_lock);
212 void ldlm_lock_touch_in_lru(struct ldlm_lock *lock)
214 struct ldlm_namespace *ns = lock->l_resource->lr_namespace;
216 spin_lock(&ns->ns_unused_lock);
217 if (!list_empty(&lock->l_lru)) {
218 ldlm_lock_remove_from_lru_nolock(lock);
219 ldlm_lock_add_to_lru_nolock(lock);
221 spin_unlock(&ns->ns_unused_lock);
225 /* This used to have a 'strict' flag, which recovery would use to mark an
226 * in-use lock as needing-to-die. Lest I am ever tempted to put it back, I
227 * shall explain why it's gone: with the new hash table scheme, once you call
228 * ldlm_lock_destroy, you can never drop your final references on this lock.
229 * Because it's not in the hash table anymore. -phil */
230 int ldlm_lock_destroy_internal(struct ldlm_lock *lock)
234 if (lock->l_readers || lock->l_writers) {
235 LDLM_ERROR(lock, "lock still has references");
236 ldlm_lock_dump(D_ERROR, lock, 0);
240 if (!list_empty(&lock->l_res_link)) {
241 LDLM_ERROR(lock, "lock still on resource");
242 ldlm_lock_dump(D_ERROR, lock, 0);
246 if (lock->l_destroyed) {
247 LASSERT(list_empty(&lock->l_lru));
251 lock->l_destroyed = 1;
254 spin_lock(&lock->l_export->exp_ldlm_data.led_lock);
255 list_del_init(&lock->l_export_chain);
257 spin_unlock(&lock->l_export->exp_ldlm_data.led_lock);
259 ldlm_lock_remove_from_lru(lock);
260 class_handle_unhash(&lock->l_handle);
263 /* Wake anyone waiting for this lock */
264 /* FIXME: I should probably add yet another flag, instead of using
265 * l_export to only call this on clients */
267 class_export_put(lock->l_export);
268 lock->l_export = NULL;
269 if (lock->l_export && lock->l_completion_ast)
270 lock->l_completion_ast(lock, 0);
276 void ldlm_lock_destroy(struct ldlm_lock *lock)
280 lock_res_and_lock(lock);
281 first = ldlm_lock_destroy_internal(lock);
282 unlock_res_and_lock(lock);
284 /* drop reference from hashtable only for first destroy */
290 void ldlm_lock_destroy_nolock(struct ldlm_lock *lock)
294 first = ldlm_lock_destroy_internal(lock);
295 /* drop reference from hashtable only for first destroy */
301 /* this is called by portals_handle2object with the handle lock taken */
302 static void lock_handle_addref(void *lock)
304 LDLM_LOCK_GET((struct ldlm_lock *)lock);
308 * usage: pass in a resource on which you have done ldlm_resource_get
309 * pass in a parent lock on which you have done a ldlm_lock_get
310 * after return, ldlm_*_put the resource and parent
311 * returns: lock with refcount 2 - one for current caller and one for remote
313 static struct ldlm_lock *ldlm_lock_new(struct ldlm_resource *resource)
315 struct ldlm_lock *lock;
318 if (resource == NULL)
321 OBD_SLAB_ALLOC(lock, ldlm_lock_slab, CFS_ALLOC_IO, sizeof(*lock));
325 spin_lock_init(&lock->l_lock);
326 lock->l_resource = ldlm_resource_getref(resource);
328 atomic_set(&lock->l_refc, 2);
329 CFS_INIT_LIST_HEAD(&lock->l_res_link);
330 CFS_INIT_LIST_HEAD(&lock->l_lru);
331 CFS_INIT_LIST_HEAD(&lock->l_export_chain);
332 CFS_INIT_LIST_HEAD(&lock->l_pending_chain);
333 CFS_INIT_LIST_HEAD(&lock->l_bl_ast);
334 CFS_INIT_LIST_HEAD(&lock->l_cp_ast);
335 cfs_waitq_init(&lock->l_waitq);
336 lock->l_blocking_lock = NULL;
337 CFS_INIT_LIST_HEAD(&lock->l_sl_mode);
338 CFS_INIT_LIST_HEAD(&lock->l_sl_policy);
340 atomic_inc(&resource->lr_namespace->ns_locks);
341 CFS_INIT_LIST_HEAD(&lock->l_handle.h_link);
342 class_handle_hash(&lock->l_handle, lock_handle_addref);
347 int ldlm_lock_change_resource(struct ldlm_namespace *ns, struct ldlm_lock *lock,
348 const struct ldlm_res_id *new_resid)
350 struct ldlm_resource *oldres = lock->l_resource;
351 struct ldlm_resource *newres;
355 LASSERT(ns_is_client(ns));
357 lock_res_and_lock(lock);
358 if (memcmp(new_resid, &lock->l_resource->lr_name,
359 sizeof(lock->l_resource->lr_name)) == 0) {
361 unlock_res_and_lock(lock);
365 LASSERT(new_resid->name[0] != 0);
367 /* This function assumes that the lock isn't on any lists */
368 LASSERT(list_empty(&lock->l_res_link));
370 type = oldres->lr_type;
371 unlock_res_and_lock(lock);
373 newres = ldlm_resource_get(ns, NULL, new_resid, type, 1);
377 lock_res_and_lock(lock);
378 LASSERT(memcmp(new_resid, &lock->l_resource->lr_name,
379 sizeof(lock->l_resource->lr_name)) != 0);
381 lock->l_resource = newres;
383 unlock_res_and_lock(lock);
385 /* ...and the flowers are still standing! */
386 ldlm_resource_putref(oldres);
395 void ldlm_lock2handle(const struct ldlm_lock *lock, struct lustre_handle *lockh)
397 lockh->cookie = lock->l_handle.h_cookie;
400 /* if flags: atomically get the lock and set the flags.
401 * Return NULL if flag already set
404 struct ldlm_lock *__ldlm_handle2lock(const struct lustre_handle *handle,
407 struct ldlm_namespace *ns;
408 struct ldlm_lock *lock, *retval = NULL;
413 lock = class_handle2object(handle->cookie);
417 LASSERT(lock->l_resource != NULL);
418 ns = lock->l_resource->lr_namespace;
421 lock_res_and_lock(lock);
423 /* It's unlikely but possible that someone marked the lock as
424 * destroyed after we did handle2object on it */
425 if (lock->l_destroyed) {
426 unlock_res_and_lock(lock);
427 CDEBUG(D_INFO, "lock already destroyed: lock %p\n", lock);
432 if (flags && (lock->l_flags & flags)) {
433 unlock_res_and_lock(lock);
439 lock->l_flags |= flags;
441 unlock_res_and_lock(lock);
448 struct ldlm_lock *ldlm_handle2lock_ns(struct ldlm_namespace *ns,
449 const struct lustre_handle *handle)
451 struct ldlm_lock *retval = NULL;
452 retval = __ldlm_handle2lock(handle, 0);
456 void ldlm_lock2desc(struct ldlm_lock *lock, struct ldlm_lock_desc *desc)
458 struct obd_export *exp = lock->l_export?:lock->l_conn_export;
459 /* INODEBITS_INTEROP: If the other side does not support
460 * inodebits, reply with a plain lock descriptor.
462 if ((lock->l_resource->lr_type == LDLM_IBITS) &&
463 (exp && !(exp->exp_connect_flags & OBD_CONNECT_IBITS))) {
464 struct ldlm_resource res = *lock->l_resource;
466 /* Make sure all the right bits are set in this lock we
467 are going to pass to client */
468 LASSERTF(lock->l_policy_data.l_inodebits.bits ==
469 (MDS_INODELOCK_LOOKUP|MDS_INODELOCK_UPDATE),
470 "Inappropriate inode lock bits during "
471 "conversion " LPU64 "\n",
472 lock->l_policy_data.l_inodebits.bits);
473 res.lr_type = LDLM_PLAIN;
474 ldlm_res2desc(&res, &desc->l_resource);
475 /* Convert "new" lock mode to something old client can
477 if ((lock->l_req_mode == LCK_CR) ||
478 (lock->l_req_mode == LCK_CW))
479 desc->l_req_mode = LCK_PR;
481 desc->l_req_mode = lock->l_req_mode;
482 if ((lock->l_granted_mode == LCK_CR) ||
483 (lock->l_granted_mode == LCK_CW)) {
484 desc->l_granted_mode = LCK_PR;
486 /* We never grant PW/EX locks to clients */
487 LASSERT((lock->l_granted_mode != LCK_PW) &&
488 (lock->l_granted_mode != LCK_EX));
489 desc->l_granted_mode = lock->l_granted_mode;
492 /* We do not copy policy here, because there is no
493 policy for plain locks */
495 ldlm_res2desc(lock->l_resource, &desc->l_resource);
496 desc->l_req_mode = lock->l_req_mode;
497 desc->l_granted_mode = lock->l_granted_mode;
498 desc->l_policy_data = lock->l_policy_data;
502 void ldlm_add_bl_work_item(struct ldlm_lock *lock, struct ldlm_lock *new,
503 struct list_head *work_list)
505 if ((lock->l_flags & LDLM_FL_AST_SENT) == 0) {
506 LDLM_DEBUG(lock, "lock incompatible; sending blocking AST.");
507 lock->l_flags |= LDLM_FL_AST_SENT;
508 /* If the enqueuing client said so, tell the AST recipient to
509 * discard dirty data, rather than writing back. */
510 if (new->l_flags & LDLM_AST_DISCARD_DATA)
511 lock->l_flags |= LDLM_FL_DISCARD_DATA;
512 LASSERT(list_empty(&lock->l_bl_ast));
513 list_add(&lock->l_bl_ast, work_list);
515 LASSERT(lock->l_blocking_lock == NULL);
516 lock->l_blocking_lock = LDLM_LOCK_GET(new);
520 void ldlm_add_cp_work_item(struct ldlm_lock *lock, struct list_head *work_list)
522 if ((lock->l_flags & LDLM_FL_CP_REQD) == 0) {
523 lock->l_flags |= LDLM_FL_CP_REQD;
524 LDLM_DEBUG(lock, "lock granted; sending completion AST.");
525 LASSERT(list_empty(&lock->l_cp_ast));
526 list_add(&lock->l_cp_ast, work_list);
531 /* must be called with lr_lock held */
532 void ldlm_add_ast_work_item(struct ldlm_lock *lock, struct ldlm_lock *new,
533 struct list_head *work_list)
536 check_res_locked(lock->l_resource);
538 ldlm_add_bl_work_item(lock, new, work_list);
540 ldlm_add_cp_work_item(lock, work_list);
544 void ldlm_lock_addref(struct lustre_handle *lockh, __u32 mode)
546 struct ldlm_lock *lock;
548 lock = ldlm_handle2lock(lockh);
549 LASSERT(lock != NULL);
550 ldlm_lock_addref_internal(lock, mode);
554 void ldlm_lock_addref_internal_nolock(struct ldlm_lock *lock, __u32 mode)
556 ldlm_lock_remove_from_lru(lock);
557 if (mode & (LCK_NL | LCK_CR | LCK_PR))
559 if (mode & (LCK_EX | LCK_CW | LCK_PW | LCK_GROUP))
562 LDLM_DEBUG(lock, "ldlm_lock_addref(%s)", ldlm_lockname[mode]);
565 /* only called for local locks */
566 void ldlm_lock_addref_internal(struct ldlm_lock *lock, __u32 mode)
568 lock_res_and_lock(lock);
569 ldlm_lock_addref_internal_nolock(lock, mode);
570 unlock_res_and_lock(lock);
573 void ldlm_lock_decref_internal(struct ldlm_lock *lock, __u32 mode)
575 struct ldlm_namespace *ns;
578 lock_res_and_lock(lock);
580 ns = lock->l_resource->lr_namespace;
582 LDLM_DEBUG(lock, "ldlm_lock_decref(%s)", ldlm_lockname[mode]);
583 if (mode & (LCK_NL | LCK_CR | LCK_PR)) {
584 LASSERT(lock->l_readers > 0);
587 if (mode & (LCK_EX | LCK_CW | LCK_PW | LCK_GROUP)) {
588 LASSERT(lock->l_writers > 0);
592 if (lock->l_flags & LDLM_FL_LOCAL &&
593 !lock->l_readers && !lock->l_writers) {
594 /* If this is a local lock on a server namespace and this was
595 * the last reference, cancel the lock. */
596 CDEBUG(D_INFO, "forcing cancel of local lock\n");
597 lock->l_flags |= LDLM_FL_CBPENDING;
600 if (!lock->l_readers && !lock->l_writers &&
601 (lock->l_flags & LDLM_FL_CBPENDING)) {
602 /* If we received a blocked AST and this was the last reference,
603 * run the callback. */
604 if (ns_is_server(ns) && lock->l_export)
605 CERROR("FL_CBPENDING set on non-local lock--just a "
608 LDLM_DEBUG(lock, "final decref done on cbpending lock");
610 LDLM_LOCK_GET(lock); /* dropped by bl thread */
611 ldlm_lock_remove_from_lru(lock);
612 unlock_res_and_lock(lock);
613 if ((lock->l_flags & LDLM_FL_ATOMIC_CB) ||
614 ldlm_bl_to_thread_lock(ns, NULL, lock) != 0)
615 ldlm_handle_bl_callback(ns, NULL, lock);
616 } else if (ns_is_client(ns) &&
617 !lock->l_readers && !lock->l_writers &&
618 !(lock->l_flags & LDLM_FL_NO_LRU) &&
619 !(lock->l_flags & LDLM_FL_BL_AST)) {
620 /* If this is a client-side namespace and this was the last
621 * reference, put it on the LRU. */
622 ldlm_lock_add_to_lru(lock);
623 unlock_res_and_lock(lock);
624 /* Call ldlm_cancel_lru() only if EARLY_CANCEL and LRU RESIZE
625 * are not supported by the server, otherwise, it is done on
627 if (!exp_connect_cancelset(lock->l_conn_export) &&
628 !ns_connect_lru_resize(ns))
629 ldlm_cancel_lru(ns, 0, LDLM_ASYNC, 0);
631 unlock_res_and_lock(lock);
634 LDLM_LOCK_PUT(lock); /* matches the ldlm_lock_get in addref */
639 void ldlm_lock_decref(struct lustre_handle *lockh, __u32 mode)
641 struct ldlm_lock *lock = __ldlm_handle2lock(lockh, 0);
642 LASSERTF(lock != NULL, "Non-existing lock: "LPX64"\n", lockh->cookie);
643 ldlm_lock_decref_internal(lock, mode);
647 /* This will drop a lock reference and mark it for destruction, but will not
648 * necessarily cancel the lock before returning. */
649 void ldlm_lock_decref_and_cancel(struct lustre_handle *lockh, __u32 mode)
651 struct ldlm_lock *lock = __ldlm_handle2lock(lockh, 0);
654 LASSERT(lock != NULL);
656 LDLM_DEBUG(lock, "ldlm_lock_decref(%s)", ldlm_lockname[mode]);
657 lock_res_and_lock(lock);
658 lock->l_flags |= LDLM_FL_CBPENDING;
659 unlock_res_and_lock(lock);
660 ldlm_lock_decref_internal(lock, mode);
664 struct sl_insert_point {
665 struct list_head *res_link;
666 struct list_head *mode_link;
667 struct list_head *policy_link;
671 * search_granted_lock
674 * Finds a position to insert the new lock.
676 * queue [input]: the granted list where search acts on;
677 * req [input]: the lock whose position to be located;
678 * prev [output]: positions within 3 lists to insert @req to
682 * - ldlm_grant_lock_with_skiplist
684 static void search_granted_lock(struct list_head *queue,
685 struct ldlm_lock *req,
686 struct sl_insert_point *prev)
688 struct list_head *tmp;
689 struct ldlm_lock *lock, *mode_end, *policy_end;
692 list_for_each(tmp, queue) {
693 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
695 mode_end = list_entry(lock->l_sl_mode.prev, struct ldlm_lock,
698 if (lock->l_req_mode != req->l_req_mode) {
699 /* jump to last lock of mode group */
700 tmp = &mode_end->l_res_link;
704 /* suitable mode group is found */
705 if (lock->l_resource->lr_type == LDLM_PLAIN) {
706 /* insert point is last lock of the mode group */
707 prev->res_link = &mode_end->l_res_link;
708 prev->mode_link = &mode_end->l_sl_mode;
709 prev->policy_link = &req->l_sl_policy;
712 } else if (lock->l_resource->lr_type == LDLM_IBITS) {
714 policy_end = list_entry(lock->l_sl_policy.prev,
718 if (lock->l_policy_data.l_inodebits.bits ==
719 req->l_policy_data.l_inodebits.bits) {
720 /* insert point is last lock of
721 * the policy group */
723 &policy_end->l_res_link;
725 &policy_end->l_sl_mode;
727 &policy_end->l_sl_policy;
732 if (policy_end == mode_end)
733 /* done with mode group */
736 /* jump to next policy group within the mode group */
737 tmp = policy_end->l_res_link.next;
738 lock = list_entry(tmp, struct ldlm_lock,
740 } /* loop over policy groups within the mode group */
742 /* insert point is last lock of the mode group,
743 * new policy group is started */
744 prev->res_link = &mode_end->l_res_link;
745 prev->mode_link = &mode_end->l_sl_mode;
746 prev->policy_link = &req->l_sl_policy;
750 LDLM_ERROR(lock, "is not LDLM_PLAIN or LDLM_IBITS lock");
755 /* insert point is last lock on the queue,
756 * new mode group and new policy group are started */
757 prev->res_link = queue->prev;
758 prev->mode_link = &req->l_sl_mode;
759 prev->policy_link = &req->l_sl_policy;
764 static void ldlm_granted_list_add_lock(struct ldlm_lock *lock,
765 struct sl_insert_point *prev)
767 struct ldlm_resource *res = lock->l_resource;
770 check_res_locked(res);
772 ldlm_resource_dump(D_OTHER, res);
773 CDEBUG(D_OTHER, "About to add this lock:\n");
774 ldlm_lock_dump(D_OTHER, lock, 0);
776 if (lock->l_destroyed) {
777 CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n");
781 LASSERT(list_empty(&lock->l_res_link));
782 LASSERT(list_empty(&lock->l_sl_mode));
783 LASSERT(list_empty(&lock->l_sl_policy));
785 list_add(&lock->l_res_link, prev->res_link);
786 list_add(&lock->l_sl_mode, prev->mode_link);
787 list_add(&lock->l_sl_policy, prev->policy_link);
792 static void ldlm_grant_lock_with_skiplist(struct ldlm_lock *lock)
794 struct sl_insert_point prev;
797 LASSERT(lock->l_req_mode == lock->l_granted_mode);
799 search_granted_lock(&lock->l_resource->lr_granted, lock, &prev);
800 ldlm_granted_list_add_lock(lock, &prev);
805 * - ldlm_lock_enqueue
806 * - ldlm_reprocess_queue
807 * - ldlm_lock_convert
809 * must be called with lr_lock held
811 void ldlm_grant_lock(struct ldlm_lock *lock, struct list_head *work_list)
813 struct ldlm_resource *res = lock->l_resource;
816 check_res_locked(res);
818 lock->l_granted_mode = lock->l_req_mode;
819 if (res->lr_type == LDLM_PLAIN || res->lr_type == LDLM_IBITS)
820 ldlm_grant_lock_with_skiplist(lock);
821 else if (res->lr_type == LDLM_EXTENT)
822 ldlm_extent_add_lock(res, lock);
824 ldlm_resource_add_lock(res, &res->lr_granted, lock);
826 if (lock->l_granted_mode < res->lr_most_restr)
827 res->lr_most_restr = lock->l_granted_mode;
829 if (work_list && lock->l_completion_ast != NULL)
830 ldlm_add_ast_work_item(lock, NULL, work_list);
832 ldlm_pool_add(&res->lr_namespace->ns_pool, lock);
836 /* returns a referenced lock or NULL. See the flag descriptions below, in the
837 * comment above ldlm_lock_match */
838 static struct ldlm_lock *search_queue(struct list_head *queue,
840 ldlm_policy_data_t *policy,
841 struct ldlm_lock *old_lock, int flags)
843 struct ldlm_lock *lock;
844 struct list_head *tmp;
846 list_for_each(tmp, queue) {
849 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
851 if (lock == old_lock)
854 /* llite sometimes wants to match locks that will be
855 * canceled when their users drop, but we allow it to match
856 * if it passes in CBPENDING and the lock still has users.
857 * this is generally only going to be used by children
858 * whose parents already hold a lock so forward progress
859 * can still happen. */
860 if (lock->l_flags & LDLM_FL_CBPENDING &&
861 !(flags & LDLM_FL_CBPENDING))
863 if (lock->l_flags & LDLM_FL_CBPENDING &&
864 lock->l_readers == 0 && lock->l_writers == 0)
867 if (!(lock->l_req_mode & *mode))
869 match = lock->l_req_mode;
871 if (lock->l_resource->lr_type == LDLM_EXTENT &&
872 (lock->l_policy_data.l_extent.start >
873 policy->l_extent.start ||
874 lock->l_policy_data.l_extent.end < policy->l_extent.end))
877 if (unlikely(match == LCK_GROUP) &&
878 lock->l_resource->lr_type == LDLM_EXTENT &&
879 lock->l_policy_data.l_extent.gid != policy->l_extent.gid)
882 /* We match if we have existing lock with same or wider set
884 if (lock->l_resource->lr_type == LDLM_IBITS &&
885 ((lock->l_policy_data.l_inodebits.bits &
886 policy->l_inodebits.bits) !=
887 policy->l_inodebits.bits))
890 if (lock->l_destroyed || (lock->l_flags & LDLM_FL_FAILED))
893 if ((flags & LDLM_FL_LOCAL_ONLY) &&
894 !(lock->l_flags & LDLM_FL_LOCAL))
897 if (flags & LDLM_FL_TEST_LOCK) {
899 ldlm_lock_touch_in_lru(lock);
901 ldlm_lock_addref_internal_nolock(lock, match);
910 void ldlm_lock_allow_match(struct ldlm_lock *lock)
912 lock_res_and_lock(lock);
913 lock->l_flags |= LDLM_FL_LVB_READY;
914 cfs_waitq_signal(&lock->l_waitq);
915 unlock_res_and_lock(lock);
918 /* Can be called in two ways:
920 * If 'ns' is NULL, then lockh describes an existing lock that we want to look
921 * for a duplicate of.
923 * Otherwise, all of the fields must be filled in, to match against.
925 * If 'flags' contains LDLM_FL_LOCAL_ONLY, then only match local locks on the
926 * server (ie, connh is NULL)
927 * If 'flags' contains LDLM_FL_BLOCK_GRANTED, then only locks on the granted
928 * list will be considered
929 * If 'flags' contains LDLM_FL_CBPENDING, then locks that have been marked
930 * to be canceled can still be matched as long as they still have reader
931 * or writer refernces
932 * If 'flags' contains LDLM_FL_TEST_LOCK, then don't actually reference a lock,
933 * just tell us if we would have matched.
935 * Returns 1 if it finds an already-existing lock that is compatible; in this
936 * case, lockh is filled in with a addref()ed lock
938 * we also check security context, if that failed we simply return 0 (to keep
939 * caller code unchanged), the context failure will be discovered by caller
942 ldlm_mode_t ldlm_lock_match(struct ldlm_namespace *ns, int flags,
943 const struct ldlm_res_id *res_id, ldlm_type_t type,
944 ldlm_policy_data_t *policy, ldlm_mode_t mode,
945 struct lustre_handle *lockh)
947 struct ldlm_resource *res;
948 struct ldlm_lock *lock, *old_lock = NULL;
953 old_lock = ldlm_handle2lock(lockh);
956 ns = old_lock->l_resource->lr_namespace;
957 res_id = &old_lock->l_resource->lr_name;
958 type = old_lock->l_resource->lr_type;
959 mode = old_lock->l_req_mode;
962 res = ldlm_resource_get(ns, NULL, res_id, type, 0);
964 LASSERT(old_lock == NULL);
970 lock = search_queue(&res->lr_granted, &mode, policy, old_lock, flags);
973 if (flags & LDLM_FL_BLOCK_GRANTED)
975 lock = search_queue(&res->lr_converting, &mode, policy, old_lock, flags);
978 lock = search_queue(&res->lr_waiting, &mode, policy, old_lock, flags);
985 ldlm_resource_putref(res);
988 ldlm_lock2handle(lock, lockh);
989 if ((flags & LDLM_FL_LVB_READY) &&
990 (!(lock->l_flags & LDLM_FL_LVB_READY))) {
991 struct l_wait_info lwi;
992 if (lock->l_completion_ast) {
993 int err = lock->l_completion_ast(lock,
994 LDLM_FL_WAIT_NOREPROC,
997 if (flags & LDLM_FL_TEST_LOCK)
1000 ldlm_lock_decref_internal(lock, mode);
1006 lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(obd_timeout), NULL,
1007 LWI_ON_SIGNAL_NOOP, NULL);
1009 /* XXX FIXME see comment on CAN_MATCH in lustre_dlm.h */
1010 l_wait_event(lock->l_waitq,
1011 (lock->l_flags & LDLM_FL_LVB_READY), &lwi);
1016 LDLM_DEBUG(lock, "matched ("LPU64" "LPU64")",
1017 (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1018 res_id->name[2] : policy->l_extent.start,
1019 (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1020 res_id->name[3] : policy->l_extent.end);
1022 /* check user's security context */
1023 if (lock->l_conn_export &&
1024 sptlrpc_import_check_ctx(
1025 class_exp2cliimp(lock->l_conn_export))) {
1026 if (!(flags & LDLM_FL_TEST_LOCK))
1027 ldlm_lock_decref_internal(lock, mode);
1031 if (flags & LDLM_FL_TEST_LOCK)
1032 LDLM_LOCK_PUT(lock);
1034 } else if (!(flags & LDLM_FL_TEST_LOCK)) {/*less verbose for test-only*/
1035 LDLM_DEBUG_NOLOCK("not matched ns %p type %u mode %u res "
1036 LPU64"/"LPU64" ("LPU64" "LPU64")", ns,
1037 type, mode, res_id->name[0], res_id->name[1],
1038 (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1039 res_id->name[2] :policy->l_extent.start,
1040 (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1041 res_id->name[3] : policy->l_extent.end);
1044 LDLM_LOCK_PUT(old_lock);
1046 return rc ? mode : 0;
1049 /* Returns a referenced lock */
1050 struct ldlm_lock *ldlm_lock_create(struct ldlm_namespace *ns,
1051 const struct ldlm_res_id *res_id,
1054 ldlm_blocking_callback blocking,
1055 ldlm_completion_callback completion,
1056 ldlm_glimpse_callback glimpse,
1057 void *data, __u32 lvb_len)
1059 struct ldlm_lock *lock;
1060 struct ldlm_resource *res;
1063 res = ldlm_resource_get(ns, NULL, res_id, type, 1);
1067 lock = ldlm_lock_new(res);
1068 ldlm_resource_putref(res);
1073 lock->l_req_mode = mode;
1074 lock->l_ast_data = data;
1075 lock->l_blocking_ast = blocking;
1076 lock->l_completion_ast = completion;
1077 lock->l_glimpse_ast = glimpse;
1078 lock->l_pid = cfs_curproc_pid();
1080 lock->l_tree_node = NULL;
1081 /* if this is the extent lock, allocate the interval tree node */
1082 if (type == LDLM_EXTENT) {
1083 if (ldlm_interval_alloc(lock) == NULL)
1088 lock->l_lvb_len = lvb_len;
1089 OBD_ALLOC(lock->l_lvb_data, lvb_len);
1090 if (lock->l_lvb_data == NULL)
1097 if (lock->l_lvb_data)
1098 OBD_FREE(lock->l_lvb_data, lvb_len);
1099 ldlm_interval_free(ldlm_interval_detach(lock));
1100 OBD_SLAB_FREE(lock, ldlm_lock_slab, sizeof(*lock));
1104 ldlm_error_t ldlm_lock_enqueue(struct ldlm_namespace *ns,
1105 struct ldlm_lock **lockp,
1106 void *cookie, int *flags)
1108 struct ldlm_lock *lock = *lockp;
1109 struct ldlm_resource *res = lock->l_resource;
1110 int local = ns_is_client(res->lr_namespace);
1111 ldlm_processing_policy policy;
1112 ldlm_error_t rc = ELDLM_OK;
1113 struct ldlm_interval *node = NULL;
1116 do_gettimeofday(&lock->l_enqueued_time);
1117 /* policies are not executed on the client or during replay */
1118 if ((*flags & (LDLM_FL_HAS_INTENT|LDLM_FL_REPLAY)) == LDLM_FL_HAS_INTENT
1119 && !local && ns->ns_policy) {
1120 rc = ns->ns_policy(ns, lockp, cookie, lock->l_req_mode, *flags,
1122 if (rc == ELDLM_LOCK_REPLACED) {
1123 /* The lock that was returned has already been granted,
1124 * and placed into lockp. If it's not the same as the
1125 * one we passed in, then destroy the old one and our
1126 * work here is done. */
1127 if (lock != *lockp) {
1128 ldlm_lock_destroy(lock);
1129 LDLM_LOCK_PUT(lock);
1131 *flags |= LDLM_FL_LOCK_CHANGED;
1133 } else if (rc != ELDLM_OK ||
1134 (rc == ELDLM_OK && (*flags & LDLM_FL_INTENT_ONLY))) {
1135 ldlm_lock_destroy(lock);
1140 /* For a replaying lock, it might be already in granted list. So
1141 * unlinking the lock will cause the interval node to be freed, we
1142 * have to allocate the interval node early otherwise we can't regrant
1143 * this lock in the future. - jay */
1144 if (!local && (*flags & LDLM_FL_REPLAY) && res->lr_type == LDLM_EXTENT)
1145 OBD_SLAB_ALLOC(node, ldlm_interval_slab, CFS_ALLOC_IO,
1148 lock_res_and_lock(lock);
1149 if (local && lock->l_req_mode == lock->l_granted_mode) {
1150 /* The server returned a blocked lock, but it was granted
1151 * before we got a chance to actually enqueue it. We don't
1152 * need to do anything else. */
1153 *flags &= ~(LDLM_FL_BLOCK_GRANTED |
1154 LDLM_FL_BLOCK_CONV | LDLM_FL_BLOCK_WAIT);
1155 GOTO(out, ELDLM_OK);
1158 ldlm_resource_unlink_lock(lock);
1159 if (res->lr_type == LDLM_EXTENT && lock->l_tree_node == NULL) {
1161 ldlm_lock_destroy_nolock(lock);
1162 GOTO(out, rc = -ENOMEM);
1165 CFS_INIT_LIST_HEAD(&node->li_group);
1166 ldlm_interval_attach(node, lock);
1170 /* Some flags from the enqueue want to make it into the AST, via the
1171 * lock's l_flags. */
1172 lock->l_flags |= *flags & LDLM_AST_DISCARD_DATA;
1174 /* This distinction between local lock trees is very important; a client
1175 * namespace only has information about locks taken by that client, and
1176 * thus doesn't have enough information to decide for itself if it can
1177 * be granted (below). In this case, we do exactly what the server
1178 * tells us to do, as dictated by the 'flags'.
1180 * We do exactly the same thing during recovery, when the server is
1181 * more or less trusting the clients not to lie.
1183 * FIXME (bug 268): Detect obvious lies by checking compatibility in
1184 * granted/converting queues. */
1186 if (*flags & LDLM_FL_BLOCK_CONV)
1187 ldlm_resource_add_lock(res, &res->lr_converting, lock);
1188 else if (*flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED))
1189 ldlm_resource_add_lock(res, &res->lr_waiting, lock);
1191 ldlm_grant_lock(lock, NULL);
1192 GOTO(out, ELDLM_OK);
1193 } else if (*flags & LDLM_FL_REPLAY) {
1194 if (*flags & LDLM_FL_BLOCK_CONV) {
1195 ldlm_resource_add_lock(res, &res->lr_converting, lock);
1196 GOTO(out, ELDLM_OK);
1197 } else if (*flags & LDLM_FL_BLOCK_WAIT) {
1198 ldlm_resource_add_lock(res, &res->lr_waiting, lock);
1199 GOTO(out, ELDLM_OK);
1200 } else if (*flags & LDLM_FL_BLOCK_GRANTED) {
1201 ldlm_grant_lock(lock, NULL);
1202 GOTO(out, ELDLM_OK);
1204 /* If no flags, fall through to normal enqueue path. */
1207 policy = ldlm_processing_policy_table[res->lr_type];
1208 policy(lock, flags, 1, &rc, NULL);
1211 unlock_res_and_lock(lock);
1213 OBD_SLAB_FREE(node, ldlm_interval_slab, sizeof(*node));
1217 /* Must be called with namespace taken: queue is waiting or converting. */
1218 int ldlm_reprocess_queue(struct ldlm_resource *res, struct list_head *queue,
1219 struct list_head *work_list)
1221 struct list_head *tmp, *pos;
1222 ldlm_processing_policy policy;
1224 int rc = LDLM_ITER_CONTINUE;
1228 check_res_locked(res);
1230 policy = ldlm_processing_policy_table[res->lr_type];
1233 list_for_each_safe(tmp, pos, queue) {
1234 struct ldlm_lock *pending;
1235 pending = list_entry(tmp, struct ldlm_lock, l_res_link);
1237 CDEBUG(D_INFO, "Reprocessing lock %p\n", pending);
1240 rc = policy(pending, &flags, 0, &err, work_list);
1241 if (rc != LDLM_ITER_CONTINUE)
1248 /* Helper function for ldlm_run_ast_work().
1250 * Send an existing rpc set specified by @arg->set and then
1251 * destroy it. Create new one if @do_create flag is set. */
1253 ldlm_send_and_maybe_create_set(struct ldlm_cb_set_arg *arg, int do_create)
1257 ptlrpc_set_wait(arg->set);
1258 if (arg->type == LDLM_BL_CALLBACK)
1259 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_GLIMPSE, 2);
1260 ptlrpc_set_destroy(arg->set);
1263 arg->set = ptlrpc_prep_set();
1269 ldlm_work_bl_ast_lock(struct list_head *tmp, struct ldlm_cb_set_arg *arg)
1271 struct ldlm_lock_desc d;
1272 struct ldlm_lock *lock = list_entry(tmp, struct ldlm_lock, l_bl_ast);
1275 /* nobody should touch l_bl_ast */
1276 lock_res_and_lock(lock);
1277 list_del_init(&lock->l_bl_ast);
1279 LASSERT(lock->l_flags & LDLM_FL_AST_SENT);
1280 LASSERT(lock->l_bl_ast_run == 0);
1281 LASSERT(lock->l_blocking_lock);
1282 lock->l_bl_ast_run++;
1283 unlock_res_and_lock(lock);
1285 ldlm_lock2desc(lock->l_blocking_lock, &d);
1287 LDLM_LOCK_PUT(lock->l_blocking_lock);
1288 lock->l_blocking_lock = NULL;
1289 lock->l_blocking_ast(lock, &d, (void *)arg,
1291 LDLM_LOCK_PUT(lock);
1297 ldlm_work_cp_ast_lock(struct list_head *tmp, struct ldlm_cb_set_arg *arg)
1299 struct ldlm_lock *lock = list_entry(tmp, struct ldlm_lock, l_cp_ast);
1300 ldlm_completion_callback completion_callback;
1304 /* It's possible to receive a completion AST before we've set
1305 * the l_completion_ast pointer: either because the AST arrived
1306 * before the reply, or simply because there's a small race
1307 * window between receiving the reply and finishing the local
1308 * enqueue. (bug 842)
1310 * This can't happen with the blocking_ast, however, because we
1311 * will never call the local blocking_ast until we drop our
1312 * reader/writer reference, which we won't do until we get the
1313 * reply and finish enqueueing. */
1315 /* nobody should touch l_cp_ast */
1316 lock_res_and_lock(lock);
1317 list_del_init(&lock->l_cp_ast);
1318 LASSERT(lock->l_flags & LDLM_FL_CP_REQD);
1319 /* save l_completion_ast since it can be changed by
1320 * mds_intent_policy(), see bug 14225 */
1321 completion_callback = lock->l_completion_ast;
1322 lock->l_flags &= ~LDLM_FL_CP_REQD;
1323 unlock_res_and_lock(lock);
1325 if (completion_callback != NULL) {
1326 completion_callback(lock, 0, (void *)arg);
1329 LDLM_LOCK_PUT(lock);
1335 ldlm_work_revoke_ast_lock(struct list_head *tmp, struct ldlm_cb_set_arg *arg)
1337 struct ldlm_lock_desc desc;
1338 struct ldlm_lock *lock = list_entry(tmp, struct ldlm_lock, l_export_chain);
1341 list_del_init(&lock->l_export_chain);
1343 /* the desc just pretend to exclusive */
1344 ldlm_lock2desc(lock, &desc);
1345 desc.l_req_mode = LCK_EX;
1346 desc.l_granted_mode = 0;
1348 lock->l_blocking_ast(lock, &desc, (void*)arg, LDLM_CB_BLOCKING);
1349 LDLM_LOCK_PUT(lock);
1354 int ldlm_run_ast_work(struct list_head *rpc_list, ldlm_desc_ast_t ast_type)
1356 struct ldlm_cb_set_arg arg;
1357 struct list_head *tmp, *pos;
1358 int (*work_ast_lock)(struct list_head *tmp, struct ldlm_cb_set_arg *arg);
1362 arg.set = ptlrpc_prep_set();
1363 atomic_set(&arg.restart, 0);
1365 case LDLM_WORK_BL_AST:
1366 arg.type = LDLM_BL_CALLBACK;
1367 work_ast_lock = ldlm_work_bl_ast_lock;
1369 case LDLM_WORK_CP_AST:
1370 arg.type = LDLM_CP_CALLBACK;
1371 work_ast_lock = ldlm_work_cp_ast_lock;
1373 case LDLM_WORK_REVOKE_AST:
1374 arg.type = LDLM_BL_CALLBACK;
1375 work_ast_lock = ldlm_work_revoke_ast_lock;
1382 list_for_each_safe(tmp, pos, rpc_list) {
1383 ast_count += work_ast_lock(tmp, &arg);
1385 /* Send the request set if it exceeds the PARALLEL_AST_LIMIT,
1386 * and create a new set for requests that remained in
1388 if (unlikely(ast_count == PARALLEL_AST_LIMIT)) {
1389 ldlm_send_and_maybe_create_set(&arg, 1);
1395 ldlm_send_and_maybe_create_set(&arg, 0);
1397 /* In case when number of ASTs is multiply of
1398 * PARALLEL_AST_LIMIT or @rpc_list was initially empty,
1399 * @arg.set must be destroyed here, otherwise we get
1400 * write memory leaking. */
1401 ptlrpc_set_destroy(arg.set);
1403 RETURN(atomic_read(&arg.restart) ? -ERESTART : 0);
1406 static int reprocess_one_queue(struct ldlm_resource *res, void *closure)
1408 ldlm_reprocess_all(res);
1409 return LDLM_ITER_CONTINUE;
1412 void ldlm_reprocess_all_ns(struct ldlm_namespace *ns)
1414 struct list_head *tmp;
1421 spin_lock(&ns->ns_hash_lock);
1422 for (i = 0; i < RES_HASH_SIZE; i++) {
1423 tmp = ns->ns_hash[i].next;
1424 while (tmp != &(ns->ns_hash[i])) {
1425 struct ldlm_resource *res =
1426 list_entry(tmp, struct ldlm_resource, lr_hash);
1428 ldlm_resource_getref(res);
1429 spin_unlock(&ns->ns_hash_lock);
1431 rc = reprocess_one_queue(res, NULL);
1433 spin_lock(&ns->ns_hash_lock);
1435 ldlm_resource_putref_locked(res);
1437 if (rc == LDLM_ITER_STOP)
1442 spin_unlock(&ns->ns_hash_lock);
1446 void ldlm_reprocess_all(struct ldlm_resource *res)
1448 CFS_LIST_HEAD(rpc_list);
1452 /* Local lock trees don't get reprocessed. */
1453 if (ns_is_client(res->lr_namespace)) {
1460 rc = ldlm_reprocess_queue(res, &res->lr_converting, &rpc_list);
1461 if (rc == LDLM_ITER_CONTINUE)
1462 ldlm_reprocess_queue(res, &res->lr_waiting, &rpc_list);
1465 rc = ldlm_run_ast_work(&rpc_list, LDLM_WORK_CP_AST);
1466 if (rc == -ERESTART) {
1467 LASSERT(list_empty(&rpc_list));
1473 void ldlm_cancel_callback(struct ldlm_lock *lock)
1475 check_res_locked(lock->l_resource);
1476 if (!(lock->l_flags & LDLM_FL_CANCEL)) {
1477 lock->l_flags |= LDLM_FL_CANCEL;
1478 if (lock->l_blocking_ast) {
1479 // l_check_no_ns_lock(ns);
1480 unlock_res_and_lock(lock);
1481 lock->l_blocking_ast(lock, NULL, lock->l_ast_data,
1483 lock_res_and_lock(lock);
1485 LDLM_DEBUG(lock, "no blocking ast");
1488 lock->l_flags |= LDLM_FL_BL_DONE;
1491 void ldlm_unlink_lock_skiplist(struct ldlm_lock *req)
1493 if (req->l_resource->lr_type != LDLM_PLAIN &&
1494 req->l_resource->lr_type != LDLM_IBITS)
1497 list_del_init(&req->l_sl_policy);
1498 list_del_init(&req->l_sl_mode);
1501 void ldlm_lock_cancel(struct ldlm_lock *lock)
1503 struct ldlm_resource *res;
1504 struct ldlm_namespace *ns;
1507 lock_res_and_lock(lock);
1509 res = lock->l_resource;
1510 ns = res->lr_namespace;
1512 /* Please do not, no matter how tempting, remove this LBUG without
1513 * talking to me first. -phik */
1514 if (lock->l_readers || lock->l_writers) {
1515 LDLM_ERROR(lock, "lock still has references");
1519 ldlm_del_waiting_lock(lock);
1521 /* Releases cancel callback. */
1522 ldlm_cancel_callback(lock);
1524 /* Yes, second time, just in case it was added again while we were
1525 running with no res lock in ldlm_cancel_callback */
1526 ldlm_del_waiting_lock(lock);
1527 ldlm_resource_unlink_lock(lock);
1528 ldlm_lock_destroy_nolock(lock);
1530 if (lock->l_granted_mode == lock->l_req_mode)
1531 ldlm_pool_del(&ns->ns_pool, lock);
1533 /* Make sure we will not be called again for same lock what is possible
1534 * if not to zero out lock->l_granted_mode */
1535 lock->l_granted_mode = 0;
1536 unlock_res_and_lock(lock);
1541 int ldlm_lock_set_data(struct lustre_handle *lockh, void *data)
1543 struct ldlm_lock *lock = ldlm_handle2lock(lockh);
1549 lock->l_ast_data = data;
1550 LDLM_LOCK_PUT(lock);
1554 void ldlm_cancel_locks_for_export(struct obd_export *exp)
1556 struct ldlm_lock *lock;
1557 struct ldlm_resource *res;
1559 spin_lock(&exp->exp_ldlm_data.led_lock);
1560 while(!list_empty(&exp->exp_ldlm_data.led_held_locks)) {
1561 lock = list_entry(exp->exp_ldlm_data.led_held_locks.next,
1562 struct ldlm_lock, l_export_chain);
1563 res = ldlm_resource_getref(lock->l_resource);
1564 LDLM_LOCK_GET(lock);
1565 spin_unlock(&exp->exp_ldlm_data.led_lock);
1567 LDLM_DEBUG(lock, "export %p", exp);
1568 ldlm_res_lvbo_update(res, NULL, 0, 1);
1570 ldlm_lock_cancel(lock);
1571 ldlm_reprocess_all(res);
1573 ldlm_resource_putref(res);
1574 LDLM_LOCK_PUT(lock);
1575 spin_lock(&exp->exp_ldlm_data.led_lock);
1577 spin_unlock(&exp->exp_ldlm_data.led_lock);
1580 struct ldlm_resource *ldlm_lock_convert(struct ldlm_lock *lock, int new_mode,
1583 CFS_LIST_HEAD(rpc_list);
1584 struct ldlm_resource *res;
1585 struct ldlm_namespace *ns;
1588 struct sl_insert_point prev;
1590 struct ldlm_interval *node;
1593 if (new_mode == lock->l_granted_mode) { // No changes? Just return.
1594 *flags |= LDLM_FL_BLOCK_GRANTED;
1595 RETURN(lock->l_resource);
1598 /* I can't check the type of lock here because the bitlock of lock
1599 * is not held here, so do the allocation blindly. -jay */
1600 OBD_SLAB_ALLOC(node, ldlm_interval_slab, CFS_ALLOC_IO, sizeof(*node));
1601 if (node == NULL) /* Actually, this causes EDEADLOCK to be returned */
1604 LASSERTF(new_mode == LCK_PW && lock->l_granted_mode == LCK_PR,
1605 "new_mode %u, granted %u\n", new_mode, lock->l_granted_mode);
1607 lock_res_and_lock(lock);
1609 res = lock->l_resource;
1610 ns = res->lr_namespace;
1612 old_mode = lock->l_req_mode;
1613 lock->l_req_mode = new_mode;
1614 if (res->lr_type == LDLM_PLAIN || res->lr_type == LDLM_IBITS) {
1615 /* remember the lock position where the lock might be
1616 * added back to the granted list later and also
1617 * remember the join mode for skiplist fixing. */
1618 prev.res_link = lock->l_res_link.prev;
1619 prev.mode_link = lock->l_sl_mode.prev;
1620 prev.policy_link = lock->l_sl_policy.prev;
1621 ldlm_resource_unlink_lock(lock);
1623 ldlm_resource_unlink_lock(lock);
1624 if (res->lr_type == LDLM_EXTENT) {
1625 /* FIXME: ugly code, I have to attach the lock to a
1626 * interval node again since perhaps it will be granted
1628 CFS_INIT_LIST_HEAD(&node->li_group);
1629 ldlm_interval_attach(node, lock);
1634 /* If this is a local resource, put it on the appropriate list. */
1635 if (ns_is_client(res->lr_namespace)) {
1636 if (*flags & (LDLM_FL_BLOCK_CONV | LDLM_FL_BLOCK_GRANTED)) {
1637 ldlm_resource_add_lock(res, &res->lr_converting, lock);
1639 /* This should never happen, because of the way the
1640 * server handles conversions. */
1641 LDLM_ERROR(lock, "Erroneous flags %d on local lock\n",
1645 ldlm_grant_lock(lock, &rpc_list);
1647 /* FIXME: completion handling not with ns_lock held ! */
1648 if (lock->l_completion_ast)
1649 lock->l_completion_ast(lock, 0, NULL);
1653 ldlm_processing_policy policy;
1654 policy = ldlm_processing_policy_table[res->lr_type];
1655 rc = policy(lock, &pflags, 0, &err, &rpc_list);
1656 if (rc == LDLM_ITER_STOP) {
1657 lock->l_req_mode = old_mode;
1658 if (res->lr_type == LDLM_EXTENT)
1659 ldlm_extent_add_lock(res, lock);
1661 ldlm_granted_list_add_lock(lock, &prev);
1665 *flags |= LDLM_FL_BLOCK_GRANTED;
1669 unlock_res_and_lock(lock);
1672 ldlm_run_ast_work(&rpc_list, LDLM_WORK_CP_AST);
1674 OBD_SLAB_FREE(node, ldlm_interval_slab, sizeof(*node));
1678 void ldlm_lock_dump(int level, struct ldlm_lock *lock, int pos)
1680 struct obd_device *obd = NULL;
1682 if (!((libcfs_debug | D_ERROR) & level))
1686 CDEBUG(level, " NULL LDLM lock\n");
1690 CDEBUG(level," -- Lock dump: %p/"LPX64" (rc: %d) (pos: %d) (pid: %d)\n",
1691 lock, lock->l_handle.h_cookie, atomic_read(&lock->l_refc),
1693 if (lock->l_conn_export != NULL)
1694 obd = lock->l_conn_export->exp_obd;
1695 if (lock->l_export && lock->l_export->exp_connection) {
1696 CDEBUG(level, " Node: NID %s (rhandle: "LPX64")\n",
1697 libcfs_nid2str(lock->l_export->exp_connection->c_peer.nid),
1698 lock->l_remote_handle.cookie);
1699 } else if (obd == NULL) {
1700 CDEBUG(level, " Node: local\n");
1702 struct obd_import *imp = obd->u.cli.cl_import;
1703 CDEBUG(level, " Node: NID %s (rhandle: "LPX64")\n",
1704 libcfs_nid2str(imp->imp_connection->c_peer.nid),
1705 lock->l_remote_handle.cookie);
1707 CDEBUG(level, " Resource: %p ("LPU64"/"LPU64"/"LPU64")\n",
1709 lock->l_resource->lr_name.name[0],
1710 lock->l_resource->lr_name.name[1],
1711 lock->l_resource->lr_name.name[2]);
1712 CDEBUG(level, " Req mode: %s, grant mode: %s, rc: %u, read: %d, "
1713 "write: %d flags: %#x\n", ldlm_lockname[lock->l_req_mode],
1714 ldlm_lockname[lock->l_granted_mode],
1715 atomic_read(&lock->l_refc), lock->l_readers, lock->l_writers,
1717 if (lock->l_resource->lr_type == LDLM_EXTENT)
1718 CDEBUG(level, " Extent: "LPU64" -> "LPU64
1719 " (req "LPU64"-"LPU64")\n",
1720 lock->l_policy_data.l_extent.start,
1721 lock->l_policy_data.l_extent.end,
1722 lock->l_req_extent.start, lock->l_req_extent.end);
1723 else if (lock->l_resource->lr_type == LDLM_FLOCK)
1724 CDEBUG(level, " Pid: %d Extent: "LPU64" -> "LPU64"\n",
1725 lock->l_policy_data.l_flock.pid,
1726 lock->l_policy_data.l_flock.start,
1727 lock->l_policy_data.l_flock.end);
1728 else if (lock->l_resource->lr_type == LDLM_IBITS)
1729 CDEBUG(level, " Bits: "LPX64"\n",
1730 lock->l_policy_data.l_inodebits.bits);
1733 void ldlm_lock_dump_handle(int level, struct lustre_handle *lockh)
1735 struct ldlm_lock *lock;
1737 if (!((libcfs_debug | D_ERROR) & level))
1740 lock = ldlm_handle2lock(lockh);
1744 ldlm_lock_dump(D_OTHER, lock, 0);
1746 LDLM_LOCK_PUT(lock);
1749 void _ldlm_lock_debug(struct ldlm_lock *lock, __u32 level,
1750 struct libcfs_debug_msg_data *data, const char *fmt,
1754 cfs_debug_limit_state_t *cdls = data->msg_cdls;
1756 va_start(args, fmt);
1758 if (lock->l_resource == NULL) {
1759 libcfs_debug_vmsg2(cdls, data->msg_subsys, level, data->msg_file,
1760 data->msg_fn, data->msg_line, fmt, args,
1761 " ns: \?\? lock: %p/"LPX64" lrc: %d/%d,%d mode: %s/%s "
1762 "res: \?\? rrc=\?\? type: \?\?\? flags: %x remote: "
1763 LPX64" expref: %d pid: %u\n", lock,
1764 lock->l_handle.h_cookie, atomic_read(&lock->l_refc),
1765 lock->l_readers, lock->l_writers,
1766 ldlm_lockname[lock->l_granted_mode],
1767 ldlm_lockname[lock->l_req_mode],
1768 lock->l_flags, lock->l_remote_handle.cookie,
1770 atomic_read(&lock->l_export->exp_refcount) : -99,
1776 switch (lock->l_resource->lr_type) {
1778 libcfs_debug_vmsg2(cdls, data->msg_subsys, level, data->msg_file,
1779 data->msg_fn, data->msg_line, fmt, args,
1780 " ns: %s lock: %p/"LPX64" lrc: %d/%d,%d mode: %s/%s "
1781 "res: "LPU64"/"LPU64" rrc: %d type: %s ["LPU64"->"LPU64
1782 "] (req "LPU64"->"LPU64") flags: %x remote: "LPX64
1783 " expref: %d pid: %u\n",
1784 lock->l_resource->lr_namespace->ns_name, lock,
1785 lock->l_handle.h_cookie, atomic_read(&lock->l_refc),
1786 lock->l_readers, lock->l_writers,
1787 ldlm_lockname[lock->l_granted_mode],
1788 ldlm_lockname[lock->l_req_mode],
1789 lock->l_resource->lr_name.name[0],
1790 lock->l_resource->lr_name.name[1],
1791 atomic_read(&lock->l_resource->lr_refcount),
1792 ldlm_typename[lock->l_resource->lr_type],
1793 lock->l_policy_data.l_extent.start,
1794 lock->l_policy_data.l_extent.end,
1795 lock->l_req_extent.start, lock->l_req_extent.end,
1796 lock->l_flags, lock->l_remote_handle.cookie,
1798 atomic_read(&lock->l_export->exp_refcount) : -99,
1803 libcfs_debug_vmsg2(cdls, data->msg_subsys, level, data->msg_file,
1804 data->msg_fn, data->msg_line, fmt, args,
1805 " ns: %s lock: %p/"LPX64" lrc: %d/%d,%d mode: %s/%s "
1806 "res: "LPU64"/"LPU64" rrc: %d type: %s pid: %d "
1807 "["LPU64"->"LPU64"] flags: %x remote: "LPX64
1808 " expref: %d pid: %u\n",
1809 lock->l_resource->lr_namespace->ns_name, lock,
1810 lock->l_handle.h_cookie, atomic_read(&lock->l_refc),
1811 lock->l_readers, lock->l_writers,
1812 ldlm_lockname[lock->l_granted_mode],
1813 ldlm_lockname[lock->l_req_mode],
1814 lock->l_resource->lr_name.name[0],
1815 lock->l_resource->lr_name.name[1],
1816 atomic_read(&lock->l_resource->lr_refcount),
1817 ldlm_typename[lock->l_resource->lr_type],
1818 lock->l_policy_data.l_flock.pid,
1819 lock->l_policy_data.l_flock.start,
1820 lock->l_policy_data.l_flock.end,
1821 lock->l_flags, lock->l_remote_handle.cookie,
1823 atomic_read(&lock->l_export->exp_refcount) : -99,
1828 libcfs_debug_vmsg2(cdls, data->msg_subsys, level, data->msg_file,
1829 data->msg_fn, data->msg_line, fmt, args,
1830 " ns: %s lock: %p/"LPX64" lrc: %d/%d,%d mode: %s/%s "
1831 "res: "LPU64"/"LPU64" bits "LPX64" rrc: %d type: %s "
1832 "flags: %x remote: "LPX64" expref: %d "
1834 lock->l_resource->lr_namespace->ns_name,
1835 lock, lock->l_handle.h_cookie,
1836 atomic_read (&lock->l_refc),
1837 lock->l_readers, lock->l_writers,
1838 ldlm_lockname[lock->l_granted_mode],
1839 ldlm_lockname[lock->l_req_mode],
1840 lock->l_resource->lr_name.name[0],
1841 lock->l_resource->lr_name.name[1],
1842 lock->l_policy_data.l_inodebits.bits,
1843 atomic_read(&lock->l_resource->lr_refcount),
1844 ldlm_typename[lock->l_resource->lr_type],
1845 lock->l_flags, lock->l_remote_handle.cookie,
1847 atomic_read(&lock->l_export->exp_refcount) : -99,
1852 libcfs_debug_vmsg2(cdls, data->msg_subsys, level, data->msg_file,
1853 data->msg_fn, data->msg_line, fmt, args,
1854 " ns: %s lock: %p/"LPX64" lrc: %d/%d,%d mode: %s/%s "
1855 "res: "LPU64"/"LPU64" rrc: %d type: %s flags: %x "
1856 "remote: "LPX64" expref: %d pid: %u\n",
1857 lock->l_resource->lr_namespace->ns_name,
1858 lock, lock->l_handle.h_cookie,
1859 atomic_read (&lock->l_refc),
1860 lock->l_readers, lock->l_writers,
1861 ldlm_lockname[lock->l_granted_mode],
1862 ldlm_lockname[lock->l_req_mode],
1863 lock->l_resource->lr_name.name[0],
1864 lock->l_resource->lr_name.name[1],
1865 atomic_read(&lock->l_resource->lr_refcount),
1866 ldlm_typename[lock->l_resource->lr_type],
1867 lock->l_flags, lock->l_remote_handle.cookie,
1869 atomic_read(&lock->l_export->exp_refcount) : -99,
1875 EXPORT_SYMBOL(_ldlm_lock_debug);