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.sun.com/software/products/lustre/docs/GPLv2.pdf
20 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21 * CA 95054 USA or visit www.sun.com if you need additional information or
27 * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
30 * Copyright (c) 2011, 2012, Whamcloud, Inc.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
36 * lustre/ldlm/ldlm_lock.c
38 * Author: Peter Braam <braam@clusterfs.com>
39 * Author: Phil Schwan <phil@clusterfs.com>
42 #define DEBUG_SUBSYSTEM S_LDLM
45 # include <libcfs/libcfs.h>
46 # include <linux/lustre_intent.h>
48 # include <liblustre.h>
51 #include <obd_class.h>
52 #include "ldlm_internal.h"
55 char *ldlm_lockname[] = {
67 char *ldlm_typename[] = {
74 static ldlm_policy_wire_to_local_t ldlm_policy_wire18_to_local[] = {
75 [LDLM_PLAIN - LDLM_MIN_TYPE] ldlm_plain_policy_wire_to_local,
76 [LDLM_EXTENT - LDLM_MIN_TYPE] ldlm_extent_policy_wire_to_local,
77 [LDLM_FLOCK - LDLM_MIN_TYPE] ldlm_flock_policy_wire18_to_local,
78 [LDLM_IBITS - LDLM_MIN_TYPE] ldlm_ibits_policy_wire_to_local,
81 static ldlm_policy_wire_to_local_t ldlm_policy_wire21_to_local[] = {
82 [LDLM_PLAIN - LDLM_MIN_TYPE] ldlm_plain_policy_wire_to_local,
83 [LDLM_EXTENT - LDLM_MIN_TYPE] ldlm_extent_policy_wire_to_local,
84 [LDLM_FLOCK - LDLM_MIN_TYPE] ldlm_flock_policy_wire21_to_local,
85 [LDLM_IBITS - LDLM_MIN_TYPE] ldlm_ibits_policy_wire_to_local,
88 static ldlm_policy_local_to_wire_t ldlm_policy_local_to_wire[] = {
89 [LDLM_PLAIN - LDLM_MIN_TYPE] ldlm_plain_policy_local_to_wire,
90 [LDLM_EXTENT - LDLM_MIN_TYPE] ldlm_extent_policy_local_to_wire,
91 [LDLM_FLOCK - LDLM_MIN_TYPE] ldlm_flock_policy_local_to_wire,
92 [LDLM_IBITS - LDLM_MIN_TYPE] ldlm_ibits_policy_local_to_wire,
96 * Converts lock policy from local format to on the wire lock_desc format
98 void ldlm_convert_policy_to_wire(ldlm_type_t type,
99 const ldlm_policy_data_t *lpolicy,
100 ldlm_wire_policy_data_t *wpolicy)
102 ldlm_policy_local_to_wire_t convert;
104 convert = ldlm_policy_local_to_wire[type - LDLM_MIN_TYPE];
106 convert(lpolicy, wpolicy);
110 * Converts lock policy from on the wire lock_desc format to local format
112 void ldlm_convert_policy_to_local(struct obd_export *exp, ldlm_type_t type,
113 const ldlm_wire_policy_data_t *wpolicy,
114 ldlm_policy_data_t *lpolicy)
116 ldlm_policy_wire_to_local_t convert;
119 /** some badnes for 2.0.0 clients, but 2.0.0 isn't supported */
120 new_client = (exp->exp_connect_flags & OBD_CONNECT_FULL20) != 0;
122 convert = ldlm_policy_wire21_to_local[type - LDLM_MIN_TYPE];
124 convert = ldlm_policy_wire18_to_local[type - LDLM_MIN_TYPE];
126 convert(wpolicy, lpolicy);
129 char *ldlm_it2str(int it)
136 case (IT_OPEN | IT_CREAT):
151 CERROR("Unknown intent %d\n", it);
156 extern cfs_mem_cache_t *ldlm_lock_slab;
158 #ifdef HAVE_SERVER_SUPPORT
159 static ldlm_processing_policy ldlm_processing_policy_table[] = {
160 [LDLM_PLAIN] ldlm_process_plain_lock,
161 [LDLM_EXTENT] ldlm_process_extent_lock,
163 [LDLM_FLOCK] ldlm_process_flock_lock,
165 [LDLM_IBITS] ldlm_process_inodebits_lock,
168 ldlm_processing_policy ldlm_get_processing_policy(struct ldlm_resource *res)
170 return ldlm_processing_policy_table[res->lr_type];
172 #endif /* HAVE_SERVER_SUPPORT */
174 void ldlm_register_intent(struct ldlm_namespace *ns, ldlm_res_policy arg)
180 * REFCOUNTED LOCK OBJECTS
185 * Lock refcounts, during creation:
186 * - one special one for allocation, dec'd only once in destroy
187 * - one for being a lock that's in-use
188 * - one for the addref associated with a new lock
190 struct ldlm_lock *ldlm_lock_get(struct ldlm_lock *lock)
192 cfs_atomic_inc(&lock->l_refc);
196 void ldlm_lock_put(struct ldlm_lock *lock)
200 LASSERT(lock->l_resource != LP_POISON);
201 LASSERT(cfs_atomic_read(&lock->l_refc) > 0);
202 if (cfs_atomic_dec_and_test(&lock->l_refc)) {
203 struct ldlm_resource *res;
206 "final lock_put on destroyed lock, freeing it.");
208 res = lock->l_resource;
209 LASSERT(lock->l_destroyed);
210 LASSERT(cfs_list_empty(&lock->l_res_link));
211 LASSERT(cfs_list_empty(&lock->l_pending_chain));
213 lprocfs_counter_decr(ldlm_res_to_ns(res)->ns_stats,
215 lu_ref_del(&res->lr_reference, "lock", lock);
216 ldlm_resource_putref(res);
217 lock->l_resource = NULL;
218 if (lock->l_export) {
219 class_export_lock_put(lock->l_export, lock);
220 lock->l_export = NULL;
223 if (lock->l_lvb_data != NULL)
224 OBD_FREE(lock->l_lvb_data, lock->l_lvb_len);
226 ldlm_interval_free(ldlm_interval_detach(lock));
227 lu_ref_fini(&lock->l_reference);
228 OBD_FREE_RCU(lock, sizeof(*lock), &lock->l_handle);
234 int ldlm_lock_remove_from_lru_nolock(struct ldlm_lock *lock)
237 if (!cfs_list_empty(&lock->l_lru)) {
238 struct ldlm_namespace *ns = ldlm_lock_to_ns(lock);
240 LASSERT(lock->l_resource->lr_type != LDLM_FLOCK);
241 cfs_list_del_init(&lock->l_lru);
242 if (lock->l_flags & LDLM_FL_SKIPPED)
243 lock->l_flags &= ~LDLM_FL_SKIPPED;
244 LASSERT(ns->ns_nr_unused > 0);
251 int ldlm_lock_remove_from_lru(struct ldlm_lock *lock)
253 struct ldlm_namespace *ns = ldlm_lock_to_ns(lock);
257 if (lock->l_ns_srv) {
258 LASSERT(cfs_list_empty(&lock->l_lru));
262 cfs_spin_lock(&ns->ns_lock);
263 rc = ldlm_lock_remove_from_lru_nolock(lock);
264 cfs_spin_unlock(&ns->ns_lock);
269 void ldlm_lock_add_to_lru_nolock(struct ldlm_lock *lock)
271 struct ldlm_namespace *ns = ldlm_lock_to_ns(lock);
273 lock->l_last_used = cfs_time_current();
274 LASSERT(cfs_list_empty(&lock->l_lru));
275 LASSERT(lock->l_resource->lr_type != LDLM_FLOCK);
276 cfs_list_add_tail(&lock->l_lru, &ns->ns_unused_list);
277 LASSERT(ns->ns_nr_unused >= 0);
281 void ldlm_lock_add_to_lru(struct ldlm_lock *lock)
283 struct ldlm_namespace *ns = ldlm_lock_to_ns(lock);
286 cfs_spin_lock(&ns->ns_lock);
287 ldlm_lock_add_to_lru_nolock(lock);
288 cfs_spin_unlock(&ns->ns_lock);
292 void ldlm_lock_touch_in_lru(struct ldlm_lock *lock)
294 struct ldlm_namespace *ns = ldlm_lock_to_ns(lock);
297 if (lock->l_ns_srv) {
298 LASSERT(cfs_list_empty(&lock->l_lru));
303 cfs_spin_lock(&ns->ns_lock);
304 if (!cfs_list_empty(&lock->l_lru)) {
305 ldlm_lock_remove_from_lru_nolock(lock);
306 ldlm_lock_add_to_lru_nolock(lock);
308 cfs_spin_unlock(&ns->ns_lock);
312 /* This used to have a 'strict' flag, which recovery would use to mark an
313 * in-use lock as needing-to-die. Lest I am ever tempted to put it back, I
314 * shall explain why it's gone: with the new hash table scheme, once you call
315 * ldlm_lock_destroy, you can never drop your final references on this lock.
316 * Because it's not in the hash table anymore. -phil */
317 int ldlm_lock_destroy_internal(struct ldlm_lock *lock)
321 if (lock->l_readers || lock->l_writers) {
322 LDLM_ERROR(lock, "lock still has references");
326 if (!cfs_list_empty(&lock->l_res_link)) {
327 LDLM_ERROR(lock, "lock still on resource");
331 if (lock->l_destroyed) {
332 LASSERT(cfs_list_empty(&lock->l_lru));
336 lock->l_destroyed = 1;
338 if (lock->l_export && lock->l_export->exp_lock_hash) {
339 /* NB: it's safe to call cfs_hash_del() even lock isn't
340 * in exp_lock_hash. */
341 /* In the function below, .hs_keycmp resolves to
342 * ldlm_export_lock_keycmp() */
343 /* coverity[overrun-buffer-val] */
344 cfs_hash_del(lock->l_export->exp_lock_hash,
345 &lock->l_remote_handle, &lock->l_exp_hash);
348 ldlm_lock_remove_from_lru(lock);
349 class_handle_unhash(&lock->l_handle);
352 /* Wake anyone waiting for this lock */
353 /* FIXME: I should probably add yet another flag, instead of using
354 * l_export to only call this on clients */
356 class_export_put(lock->l_export);
357 lock->l_export = NULL;
358 if (lock->l_export && lock->l_completion_ast)
359 lock->l_completion_ast(lock, 0);
365 void ldlm_lock_destroy(struct ldlm_lock *lock)
369 lock_res_and_lock(lock);
370 first = ldlm_lock_destroy_internal(lock);
371 unlock_res_and_lock(lock);
373 /* drop reference from hashtable only for first destroy */
375 lu_ref_del(&lock->l_reference, "hash", lock);
376 LDLM_LOCK_RELEASE(lock);
381 void ldlm_lock_destroy_nolock(struct ldlm_lock *lock)
385 first = ldlm_lock_destroy_internal(lock);
386 /* drop reference from hashtable only for first destroy */
388 lu_ref_del(&lock->l_reference, "hash", lock);
389 LDLM_LOCK_RELEASE(lock);
394 /* this is called by portals_handle2object with the handle lock taken */
395 static void lock_handle_addref(void *lock)
397 LDLM_LOCK_GET((struct ldlm_lock *)lock);
400 static void lock_handle_free(void *lock, int size)
402 LASSERT(size == sizeof(struct ldlm_lock));
403 OBD_SLAB_FREE(lock, ldlm_lock_slab, size);
406 struct portals_handle_ops lock_handle_ops = {
407 .hop_addref = lock_handle_addref,
408 .hop_free = lock_handle_free,
412 * usage: pass in a resource on which you have done ldlm_resource_get
413 * new lock will take over the refcount.
414 * returns: lock with refcount 2 - one for current caller and one for remote
416 static struct ldlm_lock *ldlm_lock_new(struct ldlm_resource *resource)
418 struct ldlm_lock *lock;
421 if (resource == NULL)
424 OBD_SLAB_ALLOC_PTR_GFP(lock, ldlm_lock_slab, CFS_ALLOC_IO);
428 cfs_spin_lock_init(&lock->l_lock);
429 lock->l_resource = resource;
430 lu_ref_add(&resource->lr_reference, "lock", lock);
432 cfs_atomic_set(&lock->l_refc, 2);
433 CFS_INIT_LIST_HEAD(&lock->l_res_link);
434 CFS_INIT_LIST_HEAD(&lock->l_lru);
435 CFS_INIT_LIST_HEAD(&lock->l_pending_chain);
436 CFS_INIT_LIST_HEAD(&lock->l_bl_ast);
437 CFS_INIT_LIST_HEAD(&lock->l_cp_ast);
438 CFS_INIT_LIST_HEAD(&lock->l_rk_ast);
439 cfs_waitq_init(&lock->l_waitq);
440 lock->l_blocking_lock = NULL;
441 CFS_INIT_LIST_HEAD(&lock->l_sl_mode);
442 CFS_INIT_LIST_HEAD(&lock->l_sl_policy);
443 CFS_INIT_HLIST_NODE(&lock->l_exp_hash);
444 CFS_INIT_HLIST_NODE(&lock->l_exp_flock_hash);
446 lprocfs_counter_incr(ldlm_res_to_ns(resource)->ns_stats,
448 CFS_INIT_LIST_HEAD(&lock->l_handle.h_link);
449 class_handle_hash(&lock->l_handle, &lock_handle_ops);
451 lu_ref_init(&lock->l_reference);
452 lu_ref_add(&lock->l_reference, "hash", lock);
453 lock->l_callback_timeout = 0;
455 #if LUSTRE_TRACKS_LOCK_EXP_REFS
456 CFS_INIT_LIST_HEAD(&lock->l_exp_refs_link);
457 lock->l_exp_refs_nr = 0;
458 lock->l_exp_refs_target = NULL;
460 CFS_INIT_LIST_HEAD(&lock->l_exp_list);
465 int ldlm_lock_change_resource(struct ldlm_namespace *ns, struct ldlm_lock *lock,
466 const struct ldlm_res_id *new_resid)
468 struct ldlm_resource *oldres = lock->l_resource;
469 struct ldlm_resource *newres;
473 LASSERT(ns_is_client(ns));
475 lock_res_and_lock(lock);
476 if (memcmp(new_resid, &lock->l_resource->lr_name,
477 sizeof(lock->l_resource->lr_name)) == 0) {
479 unlock_res_and_lock(lock);
483 LASSERT(new_resid->name[0] != 0);
485 /* This function assumes that the lock isn't on any lists */
486 LASSERT(cfs_list_empty(&lock->l_res_link));
488 type = oldres->lr_type;
489 unlock_res_and_lock(lock);
491 newres = ldlm_resource_get(ns, NULL, new_resid, type, 1);
495 lu_ref_add(&newres->lr_reference, "lock", lock);
497 * To flip the lock from the old to the new resource, lock, oldres and
498 * newres have to be locked. Resource spin-locks are nested within
499 * lock->l_lock, and are taken in the memory address order to avoid
502 cfs_spin_lock(&lock->l_lock);
503 oldres = lock->l_resource;
504 if (oldres < newres) {
506 lock_res_nested(newres, LRT_NEW);
509 lock_res_nested(oldres, LRT_NEW);
511 LASSERT(memcmp(new_resid, &oldres->lr_name,
512 sizeof oldres->lr_name) != 0);
513 lock->l_resource = newres;
515 unlock_res_and_lock(lock);
517 /* ...and the flowers are still standing! */
518 lu_ref_del(&oldres->lr_reference, "lock", lock);
519 ldlm_resource_putref(oldres);
528 void ldlm_lock2handle(const struct ldlm_lock *lock, struct lustre_handle *lockh)
530 lockh->cookie = lock->l_handle.h_cookie;
533 /* if flags: atomically get the lock and set the flags.
534 * Return NULL if flag already set
537 struct ldlm_lock *__ldlm_handle2lock(const struct lustre_handle *handle,
540 struct ldlm_lock *lock;
545 lock = class_handle2object(handle->cookie);
549 /* It's unlikely but possible that someone marked the lock as
550 * destroyed after we did handle2object on it */
551 if (flags == 0 && !lock->l_destroyed) {
552 lu_ref_add(&lock->l_reference, "handle", cfs_current());
556 lock_res_and_lock(lock);
558 LASSERT(lock->l_resource != NULL);
560 lu_ref_add_atomic(&lock->l_reference, "handle", cfs_current());
561 if (unlikely(lock->l_destroyed)) {
562 unlock_res_and_lock(lock);
563 CDEBUG(D_INFO, "lock already destroyed: lock %p\n", lock);
568 if (flags && (lock->l_flags & flags)) {
569 unlock_res_and_lock(lock);
575 lock->l_flags |= flags;
577 unlock_res_and_lock(lock);
581 void ldlm_lock2desc(struct ldlm_lock *lock, struct ldlm_lock_desc *desc)
583 struct obd_export *exp = lock->l_export?:lock->l_conn_export;
584 /* INODEBITS_INTEROP: If the other side does not support
585 * inodebits, reply with a plain lock descriptor.
587 if ((lock->l_resource->lr_type == LDLM_IBITS) &&
588 (exp && !(exp->exp_connect_flags & OBD_CONNECT_IBITS))) {
589 /* Make sure all the right bits are set in this lock we
590 are going to pass to client */
591 LASSERTF(lock->l_policy_data.l_inodebits.bits ==
592 (MDS_INODELOCK_LOOKUP | MDS_INODELOCK_UPDATE |
593 MDS_INODELOCK_LAYOUT),
594 "Inappropriate inode lock bits during "
595 "conversion " LPU64 "\n",
596 lock->l_policy_data.l_inodebits.bits);
598 ldlm_res2desc(lock->l_resource, &desc->l_resource);
599 desc->l_resource.lr_type = LDLM_PLAIN;
601 /* Convert "new" lock mode to something old client can
603 if ((lock->l_req_mode == LCK_CR) ||
604 (lock->l_req_mode == LCK_CW))
605 desc->l_req_mode = LCK_PR;
607 desc->l_req_mode = lock->l_req_mode;
608 if ((lock->l_granted_mode == LCK_CR) ||
609 (lock->l_granted_mode == LCK_CW)) {
610 desc->l_granted_mode = LCK_PR;
612 /* We never grant PW/EX locks to clients */
613 LASSERT((lock->l_granted_mode != LCK_PW) &&
614 (lock->l_granted_mode != LCK_EX));
615 desc->l_granted_mode = lock->l_granted_mode;
618 /* We do not copy policy here, because there is no
619 policy for plain locks */
621 ldlm_res2desc(lock->l_resource, &desc->l_resource);
622 desc->l_req_mode = lock->l_req_mode;
623 desc->l_granted_mode = lock->l_granted_mode;
624 ldlm_convert_policy_to_wire(lock->l_resource->lr_type,
625 &lock->l_policy_data,
626 &desc->l_policy_data);
630 void ldlm_add_bl_work_item(struct ldlm_lock *lock, struct ldlm_lock *new,
631 cfs_list_t *work_list)
633 if ((lock->l_flags & LDLM_FL_AST_SENT) == 0) {
634 LDLM_DEBUG(lock, "lock incompatible; sending blocking AST.");
635 lock->l_flags |= LDLM_FL_AST_SENT;
636 /* If the enqueuing client said so, tell the AST recipient to
637 * discard dirty data, rather than writing back. */
638 if (new->l_flags & LDLM_AST_DISCARD_DATA)
639 lock->l_flags |= LDLM_FL_DISCARD_DATA;
640 LASSERT(cfs_list_empty(&lock->l_bl_ast));
641 cfs_list_add(&lock->l_bl_ast, work_list);
643 LASSERT(lock->l_blocking_lock == NULL);
644 lock->l_blocking_lock = LDLM_LOCK_GET(new);
648 void ldlm_add_cp_work_item(struct ldlm_lock *lock, cfs_list_t *work_list)
650 if ((lock->l_flags & LDLM_FL_CP_REQD) == 0) {
651 lock->l_flags |= LDLM_FL_CP_REQD;
652 LDLM_DEBUG(lock, "lock granted; sending completion AST.");
653 LASSERT(cfs_list_empty(&lock->l_cp_ast));
654 cfs_list_add(&lock->l_cp_ast, work_list);
659 /* must be called with lr_lock held */
660 void ldlm_add_ast_work_item(struct ldlm_lock *lock, struct ldlm_lock *new,
661 cfs_list_t *work_list)
664 check_res_locked(lock->l_resource);
666 ldlm_add_bl_work_item(lock, new, work_list);
668 ldlm_add_cp_work_item(lock, work_list);
672 void ldlm_lock_addref(struct lustre_handle *lockh, __u32 mode)
674 struct ldlm_lock *lock;
676 lock = ldlm_handle2lock(lockh);
677 LASSERT(lock != NULL);
678 ldlm_lock_addref_internal(lock, mode);
682 void ldlm_lock_addref_internal_nolock(struct ldlm_lock *lock, __u32 mode)
684 ldlm_lock_remove_from_lru(lock);
685 if (mode & (LCK_NL | LCK_CR | LCK_PR)) {
687 lu_ref_add_atomic(&lock->l_reference, "reader", lock);
689 if (mode & (LCK_EX | LCK_CW | LCK_PW | LCK_GROUP | LCK_COS)) {
691 lu_ref_add_atomic(&lock->l_reference, "writer", lock);
694 lu_ref_add_atomic(&lock->l_reference, "user", lock);
695 LDLM_DEBUG(lock, "ldlm_lock_addref(%s)", ldlm_lockname[mode]);
699 * Attempts to addref a lock, and fails if lock is already LDLM_FL_CBPENDING
702 * \retval 0 success, lock was addref-ed
704 * \retval -EAGAIN lock is being canceled.
706 int ldlm_lock_addref_try(struct lustre_handle *lockh, __u32 mode)
708 struct ldlm_lock *lock;
712 lock = ldlm_handle2lock(lockh);
714 lock_res_and_lock(lock);
715 if (lock->l_readers != 0 || lock->l_writers != 0 ||
716 !(lock->l_flags & LDLM_FL_CBPENDING)) {
717 ldlm_lock_addref_internal_nolock(lock, mode);
720 unlock_res_and_lock(lock);
726 /* only called for local locks */
727 void ldlm_lock_addref_internal(struct ldlm_lock *lock, __u32 mode)
729 lock_res_and_lock(lock);
730 ldlm_lock_addref_internal_nolock(lock, mode);
731 unlock_res_and_lock(lock);
734 /* only called in ldlm_flock_destroy and for local locks.
735 * * for LDLM_FLOCK type locks, l_blocking_ast is null, and
736 * * ldlm_lock_remove_from_lru() does nothing, it is safe
737 * * for ldlm_flock_destroy usage by dropping some code */
738 void ldlm_lock_decref_internal_nolock(struct ldlm_lock *lock, __u32 mode)
740 LDLM_DEBUG(lock, "ldlm_lock_decref(%s)", ldlm_lockname[mode]);
741 if (mode & (LCK_NL | LCK_CR | LCK_PR)) {
742 LASSERT(lock->l_readers > 0);
743 lu_ref_del(&lock->l_reference, "reader", lock);
746 if (mode & (LCK_EX | LCK_CW | LCK_PW | LCK_GROUP | LCK_COS)) {
747 LASSERT(lock->l_writers > 0);
748 lu_ref_del(&lock->l_reference, "writer", lock);
752 lu_ref_del(&lock->l_reference, "user", lock);
753 LDLM_LOCK_RELEASE(lock); /* matches the LDLM_LOCK_GET() in addref */
756 void ldlm_lock_decref_internal(struct ldlm_lock *lock, __u32 mode)
758 struct ldlm_namespace *ns;
761 lock_res_and_lock(lock);
763 ns = ldlm_lock_to_ns(lock);
765 ldlm_lock_decref_internal_nolock(lock, mode);
767 if (lock->l_flags & LDLM_FL_LOCAL &&
768 !lock->l_readers && !lock->l_writers) {
769 /* If this is a local lock on a server namespace and this was
770 * the last reference, cancel the lock. */
771 CDEBUG(D_INFO, "forcing cancel of local lock\n");
772 lock->l_flags |= LDLM_FL_CBPENDING;
775 if (!lock->l_readers && !lock->l_writers &&
776 (lock->l_flags & LDLM_FL_CBPENDING)) {
777 /* If we received a blocked AST and this was the last reference,
778 * run the callback. */
779 if (lock->l_ns_srv && lock->l_export)
780 CERROR("FL_CBPENDING set on non-local lock--just a "
783 LDLM_DEBUG(lock, "final decref done on cbpending lock");
785 LDLM_LOCK_GET(lock); /* dropped by bl thread */
786 ldlm_lock_remove_from_lru(lock);
787 unlock_res_and_lock(lock);
789 if (lock->l_flags & LDLM_FL_FAIL_LOC)
790 OBD_RACE(OBD_FAIL_LDLM_CP_BL_RACE);
792 if ((lock->l_flags & LDLM_FL_ATOMIC_CB) ||
793 ldlm_bl_to_thread_lock(ns, NULL, lock) != 0)
794 ldlm_handle_bl_callback(ns, NULL, lock);
795 } else if (ns_is_client(ns) &&
796 !lock->l_readers && !lock->l_writers &&
797 !(lock->l_flags & LDLM_FL_NO_LRU) &&
798 !(lock->l_flags & LDLM_FL_BL_AST)) {
800 LDLM_DEBUG(lock, "add lock into lru list");
802 /* If this is a client-side namespace and this was the last
803 * reference, put it on the LRU. */
804 ldlm_lock_add_to_lru(lock);
805 unlock_res_and_lock(lock);
807 if (lock->l_flags & LDLM_FL_FAIL_LOC)
808 OBD_RACE(OBD_FAIL_LDLM_CP_BL_RACE);
810 /* Call ldlm_cancel_lru() only if EARLY_CANCEL and LRU RESIZE
811 * are not supported by the server, otherwise, it is done on
813 if (!exp_connect_cancelset(lock->l_conn_export) &&
814 !ns_connect_lru_resize(ns))
815 ldlm_cancel_lru(ns, 0, LDLM_ASYNC, 0);
817 LDLM_DEBUG(lock, "do not add lock into lru list");
818 unlock_res_and_lock(lock);
824 void ldlm_lock_decref(struct lustre_handle *lockh, __u32 mode)
826 struct ldlm_lock *lock = __ldlm_handle2lock(lockh, 0);
827 LASSERTF(lock != NULL, "Non-existing lock: "LPX64"\n", lockh->cookie);
828 ldlm_lock_decref_internal(lock, mode);
832 /* This will drop a lock reference and mark it for destruction, but will not
833 * necessarily cancel the lock before returning. */
834 void ldlm_lock_decref_and_cancel(struct lustre_handle *lockh, __u32 mode)
836 struct ldlm_lock *lock = __ldlm_handle2lock(lockh, 0);
839 LASSERT(lock != NULL);
841 LDLM_DEBUG(lock, "ldlm_lock_decref(%s)", ldlm_lockname[mode]);
842 lock_res_and_lock(lock);
843 lock->l_flags |= LDLM_FL_CBPENDING;
844 unlock_res_and_lock(lock);
845 ldlm_lock_decref_internal(lock, mode);
849 struct sl_insert_point {
850 cfs_list_t *res_link;
851 cfs_list_t *mode_link;
852 cfs_list_t *policy_link;
856 * search_granted_lock
859 * Finds a position to insert the new lock.
861 * queue [input]: the granted list where search acts on;
862 * req [input]: the lock whose position to be located;
863 * prev [output]: positions within 3 lists to insert @req to
867 * - ldlm_grant_lock_with_skiplist
869 static void search_granted_lock(cfs_list_t *queue,
870 struct ldlm_lock *req,
871 struct sl_insert_point *prev)
874 struct ldlm_lock *lock, *mode_end, *policy_end;
877 cfs_list_for_each(tmp, queue) {
878 lock = cfs_list_entry(tmp, struct ldlm_lock, l_res_link);
880 mode_end = cfs_list_entry(lock->l_sl_mode.prev,
881 struct ldlm_lock, l_sl_mode);
883 if (lock->l_req_mode != req->l_req_mode) {
884 /* jump to last lock of mode group */
885 tmp = &mode_end->l_res_link;
889 /* suitable mode group is found */
890 if (lock->l_resource->lr_type == LDLM_PLAIN) {
891 /* insert point is last lock of the mode group */
892 prev->res_link = &mode_end->l_res_link;
893 prev->mode_link = &mode_end->l_sl_mode;
894 prev->policy_link = &req->l_sl_policy;
897 } else if (lock->l_resource->lr_type == LDLM_IBITS) {
900 cfs_list_entry(lock->l_sl_policy.prev,
904 if (lock->l_policy_data.l_inodebits.bits ==
905 req->l_policy_data.l_inodebits.bits) {
906 /* insert point is last lock of
907 * the policy group */
909 &policy_end->l_res_link;
911 &policy_end->l_sl_mode;
913 &policy_end->l_sl_policy;
918 if (policy_end == mode_end)
919 /* done with mode group */
922 /* go to next policy group within mode group */
923 tmp = policy_end->l_res_link.next;
924 lock = cfs_list_entry(tmp, struct ldlm_lock,
926 } /* loop over policy groups within the mode group */
928 /* insert point is last lock of the mode group,
929 * new policy group is started */
930 prev->res_link = &mode_end->l_res_link;
931 prev->mode_link = &mode_end->l_sl_mode;
932 prev->policy_link = &req->l_sl_policy;
936 LDLM_ERROR(lock,"is not LDLM_PLAIN or LDLM_IBITS lock");
941 /* insert point is last lock on the queue,
942 * new mode group and new policy group are started */
943 prev->res_link = queue->prev;
944 prev->mode_link = &req->l_sl_mode;
945 prev->policy_link = &req->l_sl_policy;
950 static void ldlm_granted_list_add_lock(struct ldlm_lock *lock,
951 struct sl_insert_point *prev)
953 struct ldlm_resource *res = lock->l_resource;
956 check_res_locked(res);
958 ldlm_resource_dump(D_INFO, res);
959 LDLM_DEBUG(lock, "About to add lock:");
961 if (lock->l_destroyed) {
962 CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n");
966 LASSERT(cfs_list_empty(&lock->l_res_link));
967 LASSERT(cfs_list_empty(&lock->l_sl_mode));
968 LASSERT(cfs_list_empty(&lock->l_sl_policy));
970 cfs_list_add(&lock->l_res_link, prev->res_link);
971 cfs_list_add(&lock->l_sl_mode, prev->mode_link);
972 cfs_list_add(&lock->l_sl_policy, prev->policy_link);
977 static void ldlm_grant_lock_with_skiplist(struct ldlm_lock *lock)
979 struct sl_insert_point prev;
982 LASSERT(lock->l_req_mode == lock->l_granted_mode);
984 search_granted_lock(&lock->l_resource->lr_granted, lock, &prev);
985 ldlm_granted_list_add_lock(lock, &prev);
990 * - ldlm_lock_enqueue
991 * - ldlm_reprocess_queue
992 * - ldlm_lock_convert
994 * must be called with lr_lock held
996 void ldlm_grant_lock(struct ldlm_lock *lock, cfs_list_t *work_list)
998 struct ldlm_resource *res = lock->l_resource;
1001 check_res_locked(res);
1003 lock->l_granted_mode = lock->l_req_mode;
1004 if (res->lr_type == LDLM_PLAIN || res->lr_type == LDLM_IBITS)
1005 ldlm_grant_lock_with_skiplist(lock);
1006 else if (res->lr_type == LDLM_EXTENT)
1007 ldlm_extent_add_lock(res, lock);
1009 ldlm_resource_add_lock(res, &res->lr_granted, lock);
1011 if (lock->l_granted_mode < res->lr_most_restr)
1012 res->lr_most_restr = lock->l_granted_mode;
1014 if (work_list && lock->l_completion_ast != NULL)
1015 ldlm_add_ast_work_item(lock, NULL, work_list);
1017 ldlm_pool_add(&ldlm_res_to_ns(res)->ns_pool, lock);
1021 /* returns a referenced lock or NULL. See the flag descriptions below, in the
1022 * comment above ldlm_lock_match */
1023 static struct ldlm_lock *search_queue(cfs_list_t *queue,
1025 ldlm_policy_data_t *policy,
1026 struct ldlm_lock *old_lock,
1027 int flags, int unref)
1029 struct ldlm_lock *lock;
1032 cfs_list_for_each(tmp, queue) {
1035 lock = cfs_list_entry(tmp, struct ldlm_lock, l_res_link);
1037 if (lock == old_lock)
1040 /* llite sometimes wants to match locks that will be
1041 * canceled when their users drop, but we allow it to match
1042 * if it passes in CBPENDING and the lock still has users.
1043 * this is generally only going to be used by children
1044 * whose parents already hold a lock so forward progress
1045 * can still happen. */
1046 if (lock->l_flags & LDLM_FL_CBPENDING &&
1047 !(flags & LDLM_FL_CBPENDING))
1049 if (!unref && lock->l_flags & LDLM_FL_CBPENDING &&
1050 lock->l_readers == 0 && lock->l_writers == 0)
1053 if (!(lock->l_req_mode & *mode))
1055 match = lock->l_req_mode;
1057 if (lock->l_resource->lr_type == LDLM_EXTENT &&
1058 (lock->l_policy_data.l_extent.start >
1059 policy->l_extent.start ||
1060 lock->l_policy_data.l_extent.end < policy->l_extent.end))
1063 if (unlikely(match == LCK_GROUP) &&
1064 lock->l_resource->lr_type == LDLM_EXTENT &&
1065 lock->l_policy_data.l_extent.gid != policy->l_extent.gid)
1068 /* We match if we have existing lock with same or wider set
1070 if (lock->l_resource->lr_type == LDLM_IBITS &&
1071 ((lock->l_policy_data.l_inodebits.bits &
1072 policy->l_inodebits.bits) !=
1073 policy->l_inodebits.bits))
1077 (lock->l_destroyed || lock->l_flags & LDLM_FL_FAILED ||
1081 if ((flags & LDLM_FL_LOCAL_ONLY) &&
1082 !(lock->l_flags & LDLM_FL_LOCAL))
1085 if (flags & LDLM_FL_TEST_LOCK) {
1086 LDLM_LOCK_GET(lock);
1087 ldlm_lock_touch_in_lru(lock);
1089 ldlm_lock_addref_internal_nolock(lock, match);
1098 void ldlm_lock_fail_match_locked(struct ldlm_lock *lock)
1100 if (!lock->l_failed) {
1102 cfs_waitq_broadcast(&lock->l_waitq);
1105 EXPORT_SYMBOL(ldlm_lock_fail_match_locked);
1107 void ldlm_lock_fail_match(struct ldlm_lock *lock)
1109 lock_res_and_lock(lock);
1110 ldlm_lock_fail_match_locked(lock);
1111 unlock_res_and_lock(lock);
1113 EXPORT_SYMBOL(ldlm_lock_fail_match);
1115 void ldlm_lock_allow_match_locked(struct ldlm_lock *lock)
1117 lock->l_flags |= LDLM_FL_LVB_READY;
1118 cfs_waitq_broadcast(&lock->l_waitq);
1121 void ldlm_lock_allow_match(struct ldlm_lock *lock)
1123 lock_res_and_lock(lock);
1124 ldlm_lock_allow_match_locked(lock);
1125 unlock_res_and_lock(lock);
1128 /* Can be called in two ways:
1130 * If 'ns' is NULL, then lockh describes an existing lock that we want to look
1131 * for a duplicate of.
1133 * Otherwise, all of the fields must be filled in, to match against.
1135 * If 'flags' contains LDLM_FL_LOCAL_ONLY, then only match local locks on the
1136 * server (ie, connh is NULL)
1137 * If 'flags' contains LDLM_FL_BLOCK_GRANTED, then only locks on the granted
1138 * list will be considered
1139 * If 'flags' contains LDLM_FL_CBPENDING, then locks that have been marked
1140 * to be canceled can still be matched as long as they still have reader
1141 * or writer refernces
1142 * If 'flags' contains LDLM_FL_TEST_LOCK, then don't actually reference a lock,
1143 * just tell us if we would have matched.
1145 * Returns 1 if it finds an already-existing lock that is compatible; in this
1146 * case, lockh is filled in with a addref()ed lock
1148 * we also check security context, if that failed we simply return 0 (to keep
1149 * caller code unchanged), the context failure will be discovered by caller
1152 ldlm_mode_t ldlm_lock_match(struct ldlm_namespace *ns, int flags,
1153 const struct ldlm_res_id *res_id, ldlm_type_t type,
1154 ldlm_policy_data_t *policy, ldlm_mode_t mode,
1155 struct lustre_handle *lockh, int unref)
1157 struct ldlm_resource *res;
1158 struct ldlm_lock *lock, *old_lock = NULL;
1163 old_lock = ldlm_handle2lock(lockh);
1166 ns = ldlm_lock_to_ns(old_lock);
1167 res_id = &old_lock->l_resource->lr_name;
1168 type = old_lock->l_resource->lr_type;
1169 mode = old_lock->l_req_mode;
1172 res = ldlm_resource_get(ns, NULL, res_id, type, 0);
1174 LASSERT(old_lock == NULL);
1178 LDLM_RESOURCE_ADDREF(res);
1181 lock = search_queue(&res->lr_granted, &mode, policy, old_lock,
1185 if (flags & LDLM_FL_BLOCK_GRANTED)
1187 lock = search_queue(&res->lr_converting, &mode, policy, old_lock,
1191 lock = search_queue(&res->lr_waiting, &mode, policy, old_lock,
1199 LDLM_RESOURCE_DELREF(res);
1200 ldlm_resource_putref(res);
1203 ldlm_lock2handle(lock, lockh);
1204 if ((flags & LDLM_FL_LVB_READY) &&
1205 (!(lock->l_flags & LDLM_FL_LVB_READY))) {
1206 struct l_wait_info lwi;
1207 if (lock->l_completion_ast) {
1208 int err = lock->l_completion_ast(lock,
1209 LDLM_FL_WAIT_NOREPROC,
1212 if (flags & LDLM_FL_TEST_LOCK)
1213 LDLM_LOCK_RELEASE(lock);
1215 ldlm_lock_decref_internal(lock,
1222 lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(obd_timeout),
1223 NULL, LWI_ON_SIGNAL_NOOP, NULL);
1225 /* XXX FIXME see comment on CAN_MATCH in lustre_dlm.h */
1226 l_wait_event(lock->l_waitq,
1227 lock->l_flags & LDLM_FL_LVB_READY ||
1230 if (!(lock->l_flags & LDLM_FL_LVB_READY)) {
1231 if (flags & LDLM_FL_TEST_LOCK)
1232 LDLM_LOCK_RELEASE(lock);
1234 ldlm_lock_decref_internal(lock, mode);
1241 LDLM_DEBUG(lock, "matched ("LPU64" "LPU64")",
1242 (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1243 res_id->name[2] : policy->l_extent.start,
1244 (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1245 res_id->name[3] : policy->l_extent.end);
1247 /* check user's security context */
1248 if (lock->l_conn_export &&
1249 sptlrpc_import_check_ctx(
1250 class_exp2cliimp(lock->l_conn_export))) {
1251 if (!(flags & LDLM_FL_TEST_LOCK))
1252 ldlm_lock_decref_internal(lock, mode);
1256 if (flags & LDLM_FL_TEST_LOCK)
1257 LDLM_LOCK_RELEASE(lock);
1259 } else if (!(flags & LDLM_FL_TEST_LOCK)) {/*less verbose for test-only*/
1260 LDLM_DEBUG_NOLOCK("not matched ns %p type %u mode %u res "
1261 LPU64"/"LPU64" ("LPU64" "LPU64")", ns,
1262 type, mode, res_id->name[0], res_id->name[1],
1263 (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1264 res_id->name[2] :policy->l_extent.start,
1265 (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1266 res_id->name[3] : policy->l_extent.end);
1269 LDLM_LOCK_PUT(old_lock);
1271 return rc ? mode : 0;
1274 ldlm_mode_t ldlm_revalidate_lock_handle(struct lustre_handle *lockh,
1277 struct ldlm_lock *lock;
1278 ldlm_mode_t mode = 0;
1281 lock = ldlm_handle2lock(lockh);
1283 lock_res_and_lock(lock);
1284 if (lock->l_destroyed || lock->l_flags & LDLM_FL_FAILED ||
1288 if (lock->l_flags & LDLM_FL_CBPENDING &&
1289 lock->l_readers == 0 && lock->l_writers == 0)
1293 *bits = lock->l_policy_data.l_inodebits.bits;
1294 mode = lock->l_granted_mode;
1295 ldlm_lock_addref_internal_nolock(lock, mode);
1302 unlock_res_and_lock(lock);
1303 LDLM_LOCK_PUT(lock);
1307 EXPORT_SYMBOL(ldlm_revalidate_lock_handle);
1309 /* Returns a referenced lock */
1310 struct ldlm_lock *ldlm_lock_create(struct ldlm_namespace *ns,
1311 const struct ldlm_res_id *res_id,
1314 const struct ldlm_callback_suite *cbs,
1315 void *data, __u32 lvb_len)
1317 struct ldlm_lock *lock;
1318 struct ldlm_resource *res;
1321 res = ldlm_resource_get(ns, NULL, res_id, type, 1);
1325 lock = ldlm_lock_new(res);
1330 lock->l_req_mode = mode;
1331 lock->l_ast_data = data;
1332 lock->l_pid = cfs_curproc_pid();
1333 lock->l_ns_srv = !!ns_is_server(ns);
1335 lock->l_blocking_ast = cbs->lcs_blocking;
1336 lock->l_completion_ast = cbs->lcs_completion;
1337 lock->l_glimpse_ast = cbs->lcs_glimpse;
1338 lock->l_weigh_ast = cbs->lcs_weigh;
1341 lock->l_tree_node = NULL;
1342 /* if this is the extent lock, allocate the interval tree node */
1343 if (type == LDLM_EXTENT) {
1344 if (ldlm_interval_alloc(lock) == NULL)
1349 lock->l_lvb_len = lvb_len;
1350 OBD_ALLOC(lock->l_lvb_data, lvb_len);
1351 if (lock->l_lvb_data == NULL)
1355 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_NEW_LOCK))
1361 ldlm_lock_destroy(lock);
1362 LDLM_LOCK_RELEASE(lock);
1366 ldlm_error_t ldlm_lock_enqueue(struct ldlm_namespace *ns,
1367 struct ldlm_lock **lockp,
1368 void *cookie, int *flags)
1370 struct ldlm_lock *lock = *lockp;
1371 struct ldlm_resource *res = lock->l_resource;
1372 int local = ns_is_client(ldlm_res_to_ns(res));
1373 #ifdef HAVE_SERVER_SUPPORT
1374 ldlm_processing_policy policy;
1376 ldlm_error_t rc = ELDLM_OK;
1377 struct ldlm_interval *node = NULL;
1380 lock->l_last_activity = cfs_time_current_sec();
1381 /* policies are not executed on the client or during replay */
1382 if ((*flags & (LDLM_FL_HAS_INTENT|LDLM_FL_REPLAY)) == LDLM_FL_HAS_INTENT
1383 && !local && ns->ns_policy) {
1384 rc = ns->ns_policy(ns, lockp, cookie, lock->l_req_mode, *flags,
1386 if (rc == ELDLM_LOCK_REPLACED) {
1387 /* The lock that was returned has already been granted,
1388 * and placed into lockp. If it's not the same as the
1389 * one we passed in, then destroy the old one and our
1390 * work here is done. */
1391 if (lock != *lockp) {
1392 ldlm_lock_destroy(lock);
1393 LDLM_LOCK_RELEASE(lock);
1395 *flags |= LDLM_FL_LOCK_CHANGED;
1397 } else if (rc != ELDLM_OK ||
1398 (rc == ELDLM_OK && (*flags & LDLM_FL_INTENT_ONLY))) {
1399 ldlm_lock_destroy(lock);
1404 /* For a replaying lock, it might be already in granted list. So
1405 * unlinking the lock will cause the interval node to be freed, we
1406 * have to allocate the interval node early otherwise we can't regrant
1407 * this lock in the future. - jay */
1408 if (!local && (*flags & LDLM_FL_REPLAY) && res->lr_type == LDLM_EXTENT)
1409 OBD_SLAB_ALLOC_PTR_GFP(node, ldlm_interval_slab, CFS_ALLOC_IO);
1411 lock_res_and_lock(lock);
1412 if (local && lock->l_req_mode == lock->l_granted_mode) {
1413 /* The server returned a blocked lock, but it was granted
1414 * before we got a chance to actually enqueue it. We don't
1415 * need to do anything else. */
1416 *flags &= ~(LDLM_FL_BLOCK_GRANTED |
1417 LDLM_FL_BLOCK_CONV | LDLM_FL_BLOCK_WAIT);
1418 GOTO(out, ELDLM_OK);
1421 ldlm_resource_unlink_lock(lock);
1422 if (res->lr_type == LDLM_EXTENT && lock->l_tree_node == NULL) {
1424 ldlm_lock_destroy_nolock(lock);
1425 GOTO(out, rc = -ENOMEM);
1428 CFS_INIT_LIST_HEAD(&node->li_group);
1429 ldlm_interval_attach(node, lock);
1433 /* Some flags from the enqueue want to make it into the AST, via the
1434 * lock's l_flags. */
1435 lock->l_flags |= *flags & LDLM_AST_DISCARD_DATA;
1437 /* This distinction between local lock trees is very important; a client
1438 * namespace only has information about locks taken by that client, and
1439 * thus doesn't have enough information to decide for itself if it can
1440 * be granted (below). In this case, we do exactly what the server
1441 * tells us to do, as dictated by the 'flags'.
1443 * We do exactly the same thing during recovery, when the server is
1444 * more or less trusting the clients not to lie.
1446 * FIXME (bug 268): Detect obvious lies by checking compatibility in
1447 * granted/converting queues. */
1449 if (*flags & LDLM_FL_BLOCK_CONV)
1450 ldlm_resource_add_lock(res, &res->lr_converting, lock);
1451 else if (*flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED))
1452 ldlm_resource_add_lock(res, &res->lr_waiting, lock);
1454 ldlm_grant_lock(lock, NULL);
1455 GOTO(out, ELDLM_OK);
1456 #ifdef HAVE_SERVER_SUPPORT
1457 } else if (*flags & LDLM_FL_REPLAY) {
1458 if (*flags & LDLM_FL_BLOCK_CONV) {
1459 ldlm_resource_add_lock(res, &res->lr_converting, lock);
1460 GOTO(out, ELDLM_OK);
1461 } else if (*flags & LDLM_FL_BLOCK_WAIT) {
1462 ldlm_resource_add_lock(res, &res->lr_waiting, lock);
1463 GOTO(out, ELDLM_OK);
1464 } else if (*flags & LDLM_FL_BLOCK_GRANTED) {
1465 ldlm_grant_lock(lock, NULL);
1466 GOTO(out, ELDLM_OK);
1468 /* If no flags, fall through to normal enqueue path. */
1471 policy = ldlm_processing_policy_table[res->lr_type];
1472 policy(lock, flags, 1, &rc, NULL);
1476 CERROR("This is client-side-only module, cannot handle "
1477 "LDLM_NAMESPACE_SERVER resource type lock.\n");
1483 unlock_res_and_lock(lock);
1485 OBD_SLAB_FREE(node, ldlm_interval_slab, sizeof(*node));
1489 #ifdef HAVE_SERVER_SUPPORT
1490 /* Must be called with namespace taken: queue is waiting or converting. */
1491 int ldlm_reprocess_queue(struct ldlm_resource *res, cfs_list_t *queue,
1492 cfs_list_t *work_list)
1494 cfs_list_t *tmp, *pos;
1495 ldlm_processing_policy policy;
1497 int rc = LDLM_ITER_CONTINUE;
1501 check_res_locked(res);
1503 policy = ldlm_processing_policy_table[res->lr_type];
1506 cfs_list_for_each_safe(tmp, pos, queue) {
1507 struct ldlm_lock *pending;
1508 pending = cfs_list_entry(tmp, struct ldlm_lock, l_res_link);
1510 CDEBUG(D_INFO, "Reprocessing lock %p\n", pending);
1513 rc = policy(pending, &flags, 0, &err, work_list);
1514 if (rc != LDLM_ITER_CONTINUE)
1523 ldlm_work_bl_ast_lock(struct ptlrpc_request_set *rqset, void *opaq)
1525 struct ldlm_cb_set_arg *arg = opaq;
1526 struct ldlm_lock_desc d;
1528 struct ldlm_lock *lock;
1531 if (cfs_list_empty(arg->list))
1534 lock = cfs_list_entry(arg->list->next, struct ldlm_lock, l_bl_ast);
1536 /* nobody should touch l_bl_ast */
1537 lock_res_and_lock(lock);
1538 cfs_list_del_init(&lock->l_bl_ast);
1540 LASSERT(lock->l_flags & LDLM_FL_AST_SENT);
1541 LASSERT(lock->l_bl_ast_run == 0);
1542 LASSERT(lock->l_blocking_lock);
1543 lock->l_bl_ast_run++;
1544 unlock_res_and_lock(lock);
1546 ldlm_lock2desc(lock->l_blocking_lock, &d);
1548 rc = lock->l_blocking_ast(lock, &d, (void *)arg, LDLM_CB_BLOCKING);
1549 LDLM_LOCK_RELEASE(lock->l_blocking_lock);
1550 lock->l_blocking_lock = NULL;
1551 LDLM_LOCK_RELEASE(lock);
1557 ldlm_work_cp_ast_lock(struct ptlrpc_request_set *rqset, void *opaq)
1559 struct ldlm_cb_set_arg *arg = opaq;
1561 struct ldlm_lock *lock;
1562 ldlm_completion_callback completion_callback;
1565 if (cfs_list_empty(arg->list))
1568 lock = cfs_list_entry(arg->list->next, struct ldlm_lock, l_cp_ast);
1570 /* It's possible to receive a completion AST before we've set
1571 * the l_completion_ast pointer: either because the AST arrived
1572 * before the reply, or simply because there's a small race
1573 * window between receiving the reply and finishing the local
1574 * enqueue. (bug 842)
1576 * This can't happen with the blocking_ast, however, because we
1577 * will never call the local blocking_ast until we drop our
1578 * reader/writer reference, which we won't do until we get the
1579 * reply and finish enqueueing. */
1581 /* nobody should touch l_cp_ast */
1582 lock_res_and_lock(lock);
1583 cfs_list_del_init(&lock->l_cp_ast);
1584 LASSERT(lock->l_flags & LDLM_FL_CP_REQD);
1585 /* save l_completion_ast since it can be changed by
1586 * mds_intent_policy(), see bug 14225 */
1587 completion_callback = lock->l_completion_ast;
1588 lock->l_flags &= ~LDLM_FL_CP_REQD;
1589 unlock_res_and_lock(lock);
1591 if (completion_callback != NULL)
1592 rc = completion_callback(lock, 0, (void *)arg);
1593 LDLM_LOCK_RELEASE(lock);
1599 ldlm_work_revoke_ast_lock(struct ptlrpc_request_set *rqset, void *opaq)
1601 struct ldlm_cb_set_arg *arg = opaq;
1602 struct ldlm_lock_desc desc;
1604 struct ldlm_lock *lock;
1607 if (cfs_list_empty(arg->list))
1610 lock = cfs_list_entry(arg->list->next, struct ldlm_lock, l_rk_ast);
1611 cfs_list_del_init(&lock->l_rk_ast);
1613 /* the desc just pretend to exclusive */
1614 ldlm_lock2desc(lock, &desc);
1615 desc.l_req_mode = LCK_EX;
1616 desc.l_granted_mode = 0;
1618 rc = lock->l_blocking_ast(lock, &desc, (void*)arg, LDLM_CB_BLOCKING);
1619 LDLM_LOCK_RELEASE(lock);
1624 int ldlm_work_gl_ast_lock(struct ptlrpc_request_set *rqset, void *opaq)
1626 struct ldlm_cb_set_arg *arg = opaq;
1627 struct ldlm_glimpse_work *gl_work;
1628 struct ldlm_lock *lock;
1632 if (cfs_list_empty(arg->list))
1635 gl_work = cfs_list_entry(arg->list->next, struct ldlm_glimpse_work,
1637 cfs_list_del_init(&gl_work->gl_list);
1639 lock = gl_work->gl_lock;
1641 /* transfer the glimpse descriptor to ldlm_cb_set_arg */
1642 arg->gl_desc = gl_work->gl_desc;
1644 /* invoke the actual glimpse callback */
1645 if (lock->l_glimpse_ast(lock, (void*)arg) == 0)
1648 LDLM_LOCK_RELEASE(lock);
1650 if ((gl_work->gl_flags & LDLM_GL_WORK_NOFREE) == 0)
1651 OBD_FREE_PTR(gl_work);
1656 int ldlm_run_ast_work(struct ldlm_namespace *ns, cfs_list_t *rpc_list,
1657 ldlm_desc_ast_t ast_type)
1659 struct ldlm_cb_set_arg *arg;
1660 set_producer_func work_ast_lock;
1663 if (cfs_list_empty(rpc_list))
1670 cfs_atomic_set(&arg->restart, 0);
1671 arg->list = rpc_list;
1674 case LDLM_WORK_BL_AST:
1675 arg->type = LDLM_BL_CALLBACK;
1676 work_ast_lock = ldlm_work_bl_ast_lock;
1678 case LDLM_WORK_CP_AST:
1679 arg->type = LDLM_CP_CALLBACK;
1680 work_ast_lock = ldlm_work_cp_ast_lock;
1682 case LDLM_WORK_REVOKE_AST:
1683 arg->type = LDLM_BL_CALLBACK;
1684 work_ast_lock = ldlm_work_revoke_ast_lock;
1686 case LDLM_WORK_GL_AST:
1687 arg->type = LDLM_GL_CALLBACK;
1688 work_ast_lock = ldlm_work_gl_ast_lock;
1694 /* We create a ptlrpc request set with flow control extension.
1695 * This request set will use the work_ast_lock function to produce new
1696 * requests and will send a new request each time one completes in order
1697 * to keep the number of requests in flight to ns_max_parallel_ast */
1698 arg->set = ptlrpc_prep_fcset(ns->ns_max_parallel_ast ? : UINT_MAX,
1699 work_ast_lock, arg);
1700 if (arg->set == NULL)
1701 GOTO(out, rc = -ENOMEM);
1703 ptlrpc_set_wait(arg->set);
1704 ptlrpc_set_destroy(arg->set);
1706 rc = cfs_atomic_read(&arg->restart) ? -ERESTART : 0;
1713 static int reprocess_one_queue(struct ldlm_resource *res, void *closure)
1715 ldlm_reprocess_all(res);
1716 return LDLM_ITER_CONTINUE;
1719 static int ldlm_reprocess_res(cfs_hash_t *hs, cfs_hash_bd_t *bd,
1720 cfs_hlist_node_t *hnode, void *arg)
1722 struct ldlm_resource *res = cfs_hash_object(hs, hnode);
1725 rc = reprocess_one_queue(res, arg);
1727 return rc == LDLM_ITER_STOP;
1730 void ldlm_reprocess_all_ns(struct ldlm_namespace *ns)
1735 cfs_hash_for_each_nolock(ns->ns_rs_hash,
1736 ldlm_reprocess_res, NULL);
1741 void ldlm_reprocess_all(struct ldlm_resource *res)
1743 CFS_LIST_HEAD(rpc_list);
1745 #ifdef HAVE_SERVER_SUPPORT
1748 /* Local lock trees don't get reprocessed. */
1749 if (ns_is_client(ldlm_res_to_ns(res))) {
1756 rc = ldlm_reprocess_queue(res, &res->lr_converting, &rpc_list);
1757 if (rc == LDLM_ITER_CONTINUE)
1758 ldlm_reprocess_queue(res, &res->lr_waiting, &rpc_list);
1761 rc = ldlm_run_ast_work(ldlm_res_to_ns(res), &rpc_list,
1763 if (rc == -ERESTART) {
1764 LASSERT(cfs_list_empty(&rpc_list));
1769 if (!ns_is_client(ldlm_res_to_ns(res))) {
1770 CERROR("This is client-side-only module, cannot handle "
1771 "LDLM_NAMESPACE_SERVER resource type lock.\n");
1778 void ldlm_cancel_callback(struct ldlm_lock *lock)
1780 check_res_locked(lock->l_resource);
1781 if (!(lock->l_flags & LDLM_FL_CANCEL)) {
1782 lock->l_flags |= LDLM_FL_CANCEL;
1783 if (lock->l_blocking_ast) {
1784 // l_check_no_ns_lock(ns);
1785 unlock_res_and_lock(lock);
1786 lock->l_blocking_ast(lock, NULL, lock->l_ast_data,
1788 lock_res_and_lock(lock);
1790 LDLM_DEBUG(lock, "no blocking ast");
1793 lock->l_flags |= LDLM_FL_BL_DONE;
1796 void ldlm_unlink_lock_skiplist(struct ldlm_lock *req)
1798 if (req->l_resource->lr_type != LDLM_PLAIN &&
1799 req->l_resource->lr_type != LDLM_IBITS)
1802 cfs_list_del_init(&req->l_sl_policy);
1803 cfs_list_del_init(&req->l_sl_mode);
1806 void ldlm_lock_cancel(struct ldlm_lock *lock)
1808 struct ldlm_resource *res;
1809 struct ldlm_namespace *ns;
1812 lock_res_and_lock(lock);
1814 res = lock->l_resource;
1815 ns = ldlm_res_to_ns(res);
1817 /* Please do not, no matter how tempting, remove this LBUG without
1818 * talking to me first. -phik */
1819 if (lock->l_readers || lock->l_writers) {
1820 LDLM_ERROR(lock, "lock still has references");
1825 ldlm_del_waiting_lock(lock);
1827 /* Releases cancel callback. */
1828 ldlm_cancel_callback(lock);
1830 /* Yes, second time, just in case it was added again while we were
1831 running with no res lock in ldlm_cancel_callback */
1833 ldlm_del_waiting_lock(lock);
1835 ldlm_resource_unlink_lock(lock);
1836 ldlm_lock_destroy_nolock(lock);
1838 if (lock->l_granted_mode == lock->l_req_mode)
1839 ldlm_pool_del(&ns->ns_pool, lock);
1841 /* Make sure we will not be called again for same lock what is possible
1842 * if not to zero out lock->l_granted_mode */
1843 lock->l_granted_mode = LCK_MINMODE;
1844 unlock_res_and_lock(lock);
1849 int ldlm_lock_set_data(struct lustre_handle *lockh, void *data)
1851 struct ldlm_lock *lock = ldlm_handle2lock(lockh);
1856 if (lock->l_ast_data == NULL)
1857 lock->l_ast_data = data;
1858 if (lock->l_ast_data == data)
1860 LDLM_LOCK_PUT(lock);
1864 EXPORT_SYMBOL(ldlm_lock_set_data);
1866 struct export_cl_data {
1867 struct obd_export *ecl_exp;
1871 int ldlm_cancel_locks_for_export_cb(cfs_hash_t *hs, cfs_hash_bd_t *bd,
1872 cfs_hlist_node_t *hnode, void *data)
1875 struct export_cl_data *ecl = (struct export_cl_data *)data;
1876 struct obd_export *exp = ecl->ecl_exp;
1877 struct ldlm_lock *lock = cfs_hash_object(hs, hnode);
1878 struct ldlm_resource *res;
1880 res = ldlm_resource_getref(lock->l_resource);
1881 LDLM_LOCK_GET(lock);
1883 LDLM_DEBUG(lock, "export %p", exp);
1884 ldlm_res_lvbo_update(res, NULL, 1);
1885 ldlm_lock_cancel(lock);
1886 ldlm_reprocess_all(res);
1887 ldlm_resource_putref(res);
1888 LDLM_LOCK_RELEASE(lock);
1891 if ((ecl->ecl_loop & -ecl->ecl_loop) == ecl->ecl_loop) {
1893 "Cancel lock %p for export %p (loop %d), still have "
1894 "%d locks left on hash table.\n",
1895 lock, exp, ecl->ecl_loop,
1896 cfs_atomic_read(&hs->hs_count));
1902 void ldlm_cancel_locks_for_export(struct obd_export *exp)
1904 struct export_cl_data ecl = {
1909 cfs_hash_for_each_empty(exp->exp_lock_hash,
1910 ldlm_cancel_locks_for_export_cb, &ecl);
1914 * Downgrade an exclusive lock.
1916 * A fast variant of ldlm_lock_convert for convertion of exclusive
1917 * locks. The convertion is always successful.
1919 * \param lock A lock to convert
1920 * \param new_mode new lock mode
1922 void ldlm_lock_downgrade(struct ldlm_lock *lock, int new_mode)
1926 LASSERT(lock->l_granted_mode & (LCK_PW | LCK_EX));
1927 LASSERT(new_mode == LCK_COS);
1929 lock_res_and_lock(lock);
1930 ldlm_resource_unlink_lock(lock);
1932 * Remove the lock from pool as it will be added again in
1933 * ldlm_grant_lock() called below.
1935 ldlm_pool_del(&ldlm_lock_to_ns(lock)->ns_pool, lock);
1937 lock->l_req_mode = new_mode;
1938 ldlm_grant_lock(lock, NULL);
1939 unlock_res_and_lock(lock);
1940 ldlm_reprocess_all(lock->l_resource);
1945 struct ldlm_resource *ldlm_lock_convert(struct ldlm_lock *lock, int new_mode,
1948 CFS_LIST_HEAD(rpc_list);
1949 struct ldlm_resource *res;
1950 struct ldlm_namespace *ns;
1952 #ifdef HAVE_SERVER_SUPPORT
1954 struct sl_insert_point prev;
1956 struct ldlm_interval *node;
1959 if (new_mode == lock->l_granted_mode) { // No changes? Just return.
1960 *flags |= LDLM_FL_BLOCK_GRANTED;
1961 RETURN(lock->l_resource);
1964 /* I can't check the type of lock here because the bitlock of lock
1965 * is not held here, so do the allocation blindly. -jay */
1966 OBD_SLAB_ALLOC_PTR_GFP(node, ldlm_interval_slab, CFS_ALLOC_IO);
1967 if (node == NULL) /* Actually, this causes EDEADLOCK to be returned */
1970 LASSERTF((new_mode == LCK_PW && lock->l_granted_mode == LCK_PR),
1971 "new_mode %u, granted %u\n", new_mode, lock->l_granted_mode);
1973 lock_res_and_lock(lock);
1975 res = lock->l_resource;
1976 ns = ldlm_res_to_ns(res);
1978 #ifdef HAVE_SERVER_SUPPORT
1979 old_mode = lock->l_req_mode;
1981 lock->l_req_mode = new_mode;
1982 if (res->lr_type == LDLM_PLAIN || res->lr_type == LDLM_IBITS) {
1983 #ifdef HAVE_SERVER_SUPPORT
1984 /* remember the lock position where the lock might be
1985 * added back to the granted list later and also
1986 * remember the join mode for skiplist fixing. */
1987 prev.res_link = lock->l_res_link.prev;
1988 prev.mode_link = lock->l_sl_mode.prev;
1989 prev.policy_link = lock->l_sl_policy.prev;
1991 ldlm_resource_unlink_lock(lock);
1993 ldlm_resource_unlink_lock(lock);
1994 if (res->lr_type == LDLM_EXTENT) {
1995 /* FIXME: ugly code, I have to attach the lock to a
1996 * interval node again since perhaps it will be granted
1998 CFS_INIT_LIST_HEAD(&node->li_group);
1999 ldlm_interval_attach(node, lock);
2005 * Remove old lock from the pool before adding the lock with new
2006 * mode below in ->policy()
2008 ldlm_pool_del(&ns->ns_pool, lock);
2010 /* If this is a local resource, put it on the appropriate list. */
2011 if (ns_is_client(ldlm_res_to_ns(res))) {
2012 if (*flags & (LDLM_FL_BLOCK_CONV | LDLM_FL_BLOCK_GRANTED)) {
2013 ldlm_resource_add_lock(res, &res->lr_converting, lock);
2015 /* This should never happen, because of the way the
2016 * server handles conversions. */
2017 LDLM_ERROR(lock, "Erroneous flags %d on local lock\n",
2021 ldlm_grant_lock(lock, &rpc_list);
2023 /* FIXME: completion handling not with lr_lock held ! */
2024 if (lock->l_completion_ast)
2025 lock->l_completion_ast(lock, 0, NULL);
2027 #ifdef HAVE_SERVER_SUPPORT
2032 ldlm_processing_policy policy;
2033 policy = ldlm_processing_policy_table[res->lr_type];
2034 rc = policy(lock, &pflags, 0, &err, &rpc_list);
2035 if (rc == LDLM_ITER_STOP) {
2036 lock->l_req_mode = old_mode;
2037 if (res->lr_type == LDLM_EXTENT)
2038 ldlm_extent_add_lock(res, lock);
2040 ldlm_granted_list_add_lock(lock, &prev);
2044 *flags |= LDLM_FL_BLOCK_GRANTED;
2050 CERROR("This is client-side-only module, cannot handle "
2051 "LDLM_NAMESPACE_SERVER resource type lock.\n");
2055 unlock_res_and_lock(lock);
2058 ldlm_run_ast_work(ns, &rpc_list, LDLM_WORK_CP_AST);
2060 OBD_SLAB_FREE(node, ldlm_interval_slab, sizeof(*node));
2064 void ldlm_lock_dump_handle(int level, struct lustre_handle *lockh)
2066 struct ldlm_lock *lock;
2068 if (!((libcfs_debug | D_ERROR) & level))
2071 lock = ldlm_handle2lock(lockh);
2075 LDLM_DEBUG_LIMIT(level, lock, "###");
2077 LDLM_LOCK_PUT(lock);
2080 void _ldlm_lock_debug(struct ldlm_lock *lock,
2081 struct libcfs_debug_msg_data *msgdata,
2082 const char *fmt, ...)
2085 struct obd_export *exp = lock->l_export;
2086 struct ldlm_resource *resource = lock->l_resource;
2087 char *nid = "local";
2089 va_start(args, fmt);
2091 if (exp && exp->exp_connection) {
2092 nid = libcfs_nid2str(exp->exp_connection->c_peer.nid);
2093 } else if (exp && exp->exp_obd != NULL) {
2094 struct obd_import *imp = exp->exp_obd->u.cli.cl_import;
2095 nid = libcfs_nid2str(imp->imp_connection->c_peer.nid);
2098 if (resource == NULL) {
2099 libcfs_debug_vmsg2(msgdata, fmt, args,
2100 " ns: \?\? lock: %p/"LPX64" lrc: %d/%d,%d mode: %s/%s "
2101 "res: \?\? rrc=\?\? type: \?\?\? flags: "LPX64" nid: %s "
2102 "remote: "LPX64" expref: %d pid: %u timeout: %lu\n",
2104 lock->l_handle.h_cookie, cfs_atomic_read(&lock->l_refc),
2105 lock->l_readers, lock->l_writers,
2106 ldlm_lockname[lock->l_granted_mode],
2107 ldlm_lockname[lock->l_req_mode],
2108 lock->l_flags, nid, lock->l_remote_handle.cookie,
2109 exp ? cfs_atomic_read(&exp->exp_refcount) : -99,
2110 lock->l_pid, lock->l_callback_timeout);
2115 switch (resource->lr_type) {
2117 libcfs_debug_vmsg2(msgdata, fmt, args,
2118 " ns: %s lock: %p/"LPX64" lrc: %d/%d,%d mode: %s/%s "
2119 "res: "LPU64"/"LPU64" rrc: %d type: %s ["LPU64"->"LPU64
2120 "] (req "LPU64"->"LPU64") flags: "LPX64" nid: %s remote:"
2121 " "LPX64" expref: %d pid: %u timeout %lu\n",
2122 ldlm_lock_to_ns_name(lock), lock,
2123 lock->l_handle.h_cookie, cfs_atomic_read(&lock->l_refc),
2124 lock->l_readers, lock->l_writers,
2125 ldlm_lockname[lock->l_granted_mode],
2126 ldlm_lockname[lock->l_req_mode],
2127 resource->lr_name.name[0],
2128 resource->lr_name.name[1],
2129 cfs_atomic_read(&resource->lr_refcount),
2130 ldlm_typename[resource->lr_type],
2131 lock->l_policy_data.l_extent.start,
2132 lock->l_policy_data.l_extent.end,
2133 lock->l_req_extent.start, lock->l_req_extent.end,
2134 lock->l_flags, nid, lock->l_remote_handle.cookie,
2135 exp ? cfs_atomic_read(&exp->exp_refcount) : -99,
2136 lock->l_pid, lock->l_callback_timeout);
2140 libcfs_debug_vmsg2(msgdata, fmt, args,
2141 " ns: %s lock: %p/"LPX64" lrc: %d/%d,%d mode: %s/%s "
2142 "res: "LPU64"/"LPU64" rrc: %d type: %s pid: %d "
2143 "["LPU64"->"LPU64"] flags: "LPX64" nid: %s remote: "LPX64
2144 " expref: %d pid: %u timeout: %lu\n",
2145 ldlm_lock_to_ns_name(lock), lock,
2146 lock->l_handle.h_cookie, cfs_atomic_read(&lock->l_refc),
2147 lock->l_readers, lock->l_writers,
2148 ldlm_lockname[lock->l_granted_mode],
2149 ldlm_lockname[lock->l_req_mode],
2150 resource->lr_name.name[0],
2151 resource->lr_name.name[1],
2152 cfs_atomic_read(&resource->lr_refcount),
2153 ldlm_typename[resource->lr_type],
2154 lock->l_policy_data.l_flock.pid,
2155 lock->l_policy_data.l_flock.start,
2156 lock->l_policy_data.l_flock.end,
2157 lock->l_flags, nid, lock->l_remote_handle.cookie,
2158 exp ? cfs_atomic_read(&exp->exp_refcount) : -99,
2159 lock->l_pid, lock->l_callback_timeout);
2163 libcfs_debug_vmsg2(msgdata, fmt, args,
2164 " ns: %s lock: %p/"LPX64" lrc: %d/%d,%d mode: %s/%s "
2165 "res: "LPU64"/"LPU64" bits "LPX64" rrc: %d type: %s "
2166 "flags: "LPX64" nid: %s remote: "LPX64" expref: %d "
2167 "pid: %u timeout: %lu\n",
2168 ldlm_lock_to_ns_name(lock),
2169 lock, lock->l_handle.h_cookie,
2170 cfs_atomic_read (&lock->l_refc),
2171 lock->l_readers, lock->l_writers,
2172 ldlm_lockname[lock->l_granted_mode],
2173 ldlm_lockname[lock->l_req_mode],
2174 resource->lr_name.name[0],
2175 resource->lr_name.name[1],
2176 lock->l_policy_data.l_inodebits.bits,
2177 cfs_atomic_read(&resource->lr_refcount),
2178 ldlm_typename[resource->lr_type],
2179 lock->l_flags, nid, lock->l_remote_handle.cookie,
2180 exp ? cfs_atomic_read(&exp->exp_refcount) : -99,
2181 lock->l_pid, lock->l_callback_timeout);
2185 libcfs_debug_vmsg2(msgdata, fmt, args,
2186 " ns: %s lock: %p/"LPX64" lrc: %d/%d,%d mode: %s/%s "
2187 "res: "LPU64"/"LPU64" rrc: %d type: %s flags: "LPX64" "
2188 "nid: %s remote: "LPX64" expref: %d pid: %u timeout %lu"
2190 ldlm_lock_to_ns_name(lock),
2191 lock, lock->l_handle.h_cookie,
2192 cfs_atomic_read (&lock->l_refc),
2193 lock->l_readers, lock->l_writers,
2194 ldlm_lockname[lock->l_granted_mode],
2195 ldlm_lockname[lock->l_req_mode],
2196 resource->lr_name.name[0],
2197 resource->lr_name.name[1],
2198 cfs_atomic_read(&resource->lr_refcount),
2199 ldlm_typename[resource->lr_type],
2200 lock->l_flags, nid, lock->l_remote_handle.cookie,
2201 exp ? cfs_atomic_read(&exp->exp_refcount) : -99,
2202 lock->l_pid, lock->l_callback_timeout);
2207 EXPORT_SYMBOL(_ldlm_lock_debug);