Whamcloud - gitweb
LU-12511 ldlm: free resource when ldlm_lock_create() fails.
[fs/lustre-release.git] / lustre / ldlm / ldlm_lock.c
1 /*
2  * GPL HEADER START
3  *
4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
5  *
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.
9  *
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).
15  *
16  * You should have received a copy of the GNU General Public License
17  * version 2 along with this program; If not, see
18  * http://www.gnu.org/licenses/gpl-2.0.html
19  *
20  * GPL HEADER END
21  */
22 /*
23  * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
24  * Use is subject to license terms.
25  *
26  * Copyright (c) 2010, 2017, Intel Corporation.
27  */
28 /*
29  * This file is part of Lustre, http://www.lustre.org/
30  *
31  * lustre/ldlm/ldlm_lock.c
32  *
33  * Author: Peter Braam <braam@clusterfs.com>
34  * Author: Phil Schwan <phil@clusterfs.com>
35  */
36
37 #define DEBUG_SUBSYSTEM S_LDLM
38
39 #include <libcfs/libcfs.h>
40
41 #include <lustre_swab.h>
42 #include <obd_class.h>
43
44 #include "ldlm_internal.h"
45
46 struct kmem_cache *ldlm_glimpse_work_kmem;
47 EXPORT_SYMBOL(ldlm_glimpse_work_kmem);
48
49 /* lock types */
50 char *ldlm_lockname[] = {
51         [0] = "--",
52         [LCK_EX] = "EX",
53         [LCK_PW] = "PW",
54         [LCK_PR] = "PR",
55         [LCK_CW] = "CW",
56         [LCK_CR] = "CR",
57         [LCK_NL] = "NL",
58         [LCK_GROUP] = "GROUP",
59         [LCK_COS] = "COS"
60 };
61 EXPORT_SYMBOL(ldlm_lockname);
62
63 char *ldlm_typename[] = {
64         [LDLM_PLAIN] = "PLN",
65         [LDLM_EXTENT] = "EXT",
66         [LDLM_FLOCK] = "FLK",
67         [LDLM_IBITS] = "IBT",
68 };
69
70 static ldlm_policy_wire_to_local_t ldlm_policy_wire_to_local[] = {
71         [LDLM_PLAIN - LDLM_MIN_TYPE]  = ldlm_plain_policy_wire_to_local,
72         [LDLM_EXTENT - LDLM_MIN_TYPE] = ldlm_extent_policy_wire_to_local,
73         [LDLM_FLOCK - LDLM_MIN_TYPE]  = ldlm_flock_policy_wire_to_local,
74         [LDLM_IBITS - LDLM_MIN_TYPE]  = ldlm_ibits_policy_wire_to_local,
75 };
76
77 static ldlm_policy_local_to_wire_t ldlm_policy_local_to_wire[] = {
78         [LDLM_PLAIN - LDLM_MIN_TYPE]  = ldlm_plain_policy_local_to_wire,
79         [LDLM_EXTENT - LDLM_MIN_TYPE] = ldlm_extent_policy_local_to_wire,
80         [LDLM_FLOCK - LDLM_MIN_TYPE]  = ldlm_flock_policy_local_to_wire,
81         [LDLM_IBITS - LDLM_MIN_TYPE]  = ldlm_ibits_policy_local_to_wire,
82 };
83
84 /**
85  * Converts lock policy from local format to on the wire lock_desc format
86  */
87 void ldlm_convert_policy_to_wire(enum ldlm_type type,
88                                  const union ldlm_policy_data *lpolicy,
89                                  union ldlm_wire_policy_data *wpolicy)
90 {
91         ldlm_policy_local_to_wire_t convert;
92
93         convert = ldlm_policy_local_to_wire[type - LDLM_MIN_TYPE];
94
95         convert(lpolicy, wpolicy);
96 }
97
98 /**
99  * Converts lock policy from on the wire lock_desc format to local format
100  */
101 void ldlm_convert_policy_to_local(struct obd_export *exp, enum ldlm_type type,
102                                   const union ldlm_wire_policy_data *wpolicy,
103                                   union ldlm_policy_data *lpolicy)
104 {
105         ldlm_policy_wire_to_local_t convert;
106
107         convert = ldlm_policy_wire_to_local[type - LDLM_MIN_TYPE];
108
109         convert(wpolicy, lpolicy);
110 }
111
112 const char *ldlm_it2str(enum ldlm_intent_flags it)
113 {
114         switch (it) {
115         case IT_OPEN:
116                 return "open";
117         case IT_CREAT:
118                 return "creat";
119         case (IT_OPEN | IT_CREAT):
120                 return "open|creat";
121         case IT_READDIR:
122                 return "readdir";
123         case IT_GETATTR:
124                 return "getattr";
125         case IT_LOOKUP:
126                 return "lookup";
127         case IT_GETXATTR:
128                 return "getxattr";
129         case IT_LAYOUT:
130                 return "layout";
131         default:
132                 CERROR("Unknown intent 0x%08x\n", it);
133                 return "UNKNOWN";
134         }
135 }
136 EXPORT_SYMBOL(ldlm_it2str);
137
138 #ifdef HAVE_SERVER_SUPPORT
139 static ldlm_processing_policy ldlm_processing_policy_table[] = {
140         [LDLM_PLAIN]    = ldlm_process_plain_lock,
141         [LDLM_EXTENT]   = ldlm_process_extent_lock,
142         [LDLM_FLOCK]    = ldlm_process_flock_lock,
143         [LDLM_IBITS]    = ldlm_process_inodebits_lock,
144 };
145
146 ldlm_processing_policy ldlm_get_processing_policy(struct ldlm_resource *res)
147 {
148         return ldlm_processing_policy_table[res->lr_type];
149 }
150 EXPORT_SYMBOL(ldlm_get_processing_policy);
151
152 static ldlm_reprocessing_policy ldlm_reprocessing_policy_table[] = {
153         [LDLM_PLAIN]    = ldlm_reprocess_queue,
154         [LDLM_EXTENT]   = ldlm_reprocess_queue,
155         [LDLM_FLOCK]    = ldlm_reprocess_queue,
156         [LDLM_IBITS]    = ldlm_reprocess_inodebits_queue,
157 };
158
159 ldlm_reprocessing_policy ldlm_get_reprocessing_policy(struct ldlm_resource *res)
160 {
161         return ldlm_reprocessing_policy_table[res->lr_type];
162 }
163
164 #endif /* HAVE_SERVER_SUPPORT */
165
166 void ldlm_register_intent(struct ldlm_namespace *ns, ldlm_res_policy arg)
167 {
168         ns->ns_policy = arg;
169 }
170 EXPORT_SYMBOL(ldlm_register_intent);
171
172 /*
173  * REFCOUNTED LOCK OBJECTS
174  */
175
176
177 /**
178  * Get a reference on a lock.
179  *
180  * Lock refcounts, during creation:
181  *   - one special one for allocation, dec'd only once in destroy
182  *   - one for being a lock that's in-use
183  *   - one for the addref associated with a new lock
184  */
185 struct ldlm_lock *ldlm_lock_get(struct ldlm_lock *lock)
186 {
187         refcount_inc(&lock->l_handle.h_ref);
188         return lock;
189 }
190 EXPORT_SYMBOL(ldlm_lock_get);
191
192 static void lock_handle_free(struct rcu_head *rcu)
193 {
194         struct ldlm_lock *lock = container_of(rcu, struct ldlm_lock,
195                                               l_handle.h_rcu);
196
197         OBD_FREE_PRE(lock, sizeof(*lock), "slab-freed");
198         kmem_cache_free(ldlm_lock_slab, lock);
199 }
200
201 /**
202  * Release lock reference.
203  *
204  * Also frees the lock if it was last reference.
205  */
206 void ldlm_lock_put(struct ldlm_lock *lock)
207 {
208         ENTRY;
209
210         LASSERT(lock->l_resource != LP_POISON);
211         LASSERT(refcount_read(&lock->l_handle.h_ref) > 0);
212         if (refcount_dec_and_test(&lock->l_handle.h_ref)) {
213                 struct ldlm_resource *res;
214
215                 LDLM_DEBUG(lock,
216                            "final lock_put on destroyed lock, freeing it.");
217
218                 res = lock->l_resource;
219                 LASSERT(ldlm_is_destroyed(lock));
220                 LASSERT(list_empty(&lock->l_exp_list));
221                 LASSERT(list_empty(&lock->l_res_link));
222                 LASSERT(list_empty(&lock->l_pending_chain));
223
224                 lprocfs_counter_decr(ldlm_res_to_ns(res)->ns_stats,
225                                      LDLM_NSS_LOCKS);
226                 lu_ref_del(&res->lr_reference, "lock", lock);
227                 if (lock->l_export) {
228                         class_export_lock_put(lock->l_export, lock);
229                         lock->l_export = NULL;
230                 }
231
232                 if (lock->l_lvb_data != NULL)
233                         OBD_FREE_LARGE(lock->l_lvb_data, lock->l_lvb_len);
234
235                 if (res->lr_type == LDLM_EXTENT) {
236                         ldlm_interval_free(ldlm_interval_detach(lock));
237                 } else if (res->lr_type == LDLM_IBITS) {
238                         if (lock->l_ibits_node != NULL)
239                                 OBD_SLAB_FREE_PTR(lock->l_ibits_node,
240                                                   ldlm_inodebits_slab);
241                 }
242                 ldlm_resource_putref(res);
243                 lock->l_resource = NULL;
244                 lu_ref_fini(&lock->l_reference);
245                 call_rcu(&lock->l_handle.h_rcu, lock_handle_free);
246         }
247
248         EXIT;
249 }
250 EXPORT_SYMBOL(ldlm_lock_put);
251
252 /**
253  * Removes LDLM lock \a lock from LRU. Assumes LRU is already locked.
254  */
255 int ldlm_lock_remove_from_lru_nolock(struct ldlm_lock *lock)
256 {
257         int rc = 0;
258         if (!list_empty(&lock->l_lru)) {
259                 struct ldlm_namespace *ns = ldlm_lock_to_ns(lock);
260
261                 LASSERT(lock->l_resource->lr_type != LDLM_FLOCK);
262                 if (ns->ns_last_pos == &lock->l_lru)
263                         ns->ns_last_pos = lock->l_lru.prev;
264                 list_del_init(&lock->l_lru);
265                 LASSERT(ns->ns_nr_unused > 0);
266                 ns->ns_nr_unused--;
267                 rc = 1;
268         }
269         return rc;
270 }
271
272 /**
273  * Removes LDLM lock \a lock from LRU. Obtains the LRU lock first.
274  *
275  * If \a last_use is non-zero, it will remove the lock from LRU only if
276  * it matches lock's l_last_used.
277  *
278  * \retval 0 if \a last_use is set, the lock is not in LRU list or \a last_use
279  *           doesn't match lock's l_last_used;
280  *           otherwise, the lock hasn't been in the LRU list.
281  * \retval 1 the lock was in LRU list and removed.
282  */
283 int ldlm_lock_remove_from_lru_check(struct ldlm_lock *lock, ktime_t last_use)
284 {
285         struct ldlm_namespace *ns = ldlm_lock_to_ns(lock);
286         int rc = 0;
287
288         ENTRY;
289         if (ldlm_is_ns_srv(lock)) {
290                 LASSERT(list_empty(&lock->l_lru));
291                 RETURN(0);
292         }
293
294         spin_lock(&ns->ns_lock);
295         if (!ktime_compare(last_use, ktime_set(0, 0)) ||
296             !ktime_compare(last_use, lock->l_last_used))
297                 rc = ldlm_lock_remove_from_lru_nolock(lock);
298         spin_unlock(&ns->ns_lock);
299
300         RETURN(rc);
301 }
302
303 /**
304  * Adds LDLM lock \a lock to namespace LRU. Assumes LRU is already locked.
305  */
306 void ldlm_lock_add_to_lru_nolock(struct ldlm_lock *lock)
307 {
308         struct ldlm_namespace *ns = ldlm_lock_to_ns(lock);
309
310         lock->l_last_used = ktime_get();
311         LASSERT(list_empty(&lock->l_lru));
312         LASSERT(lock->l_resource->lr_type != LDLM_FLOCK);
313         list_add_tail(&lock->l_lru, &ns->ns_unused_list);
314         LASSERT(ns->ns_nr_unused >= 0);
315         ns->ns_nr_unused++;
316 }
317
318 /**
319  * Adds LDLM lock \a lock to namespace LRU. Obtains necessary LRU locks
320  * first.
321  */
322 void ldlm_lock_add_to_lru(struct ldlm_lock *lock)
323 {
324         struct ldlm_namespace *ns = ldlm_lock_to_ns(lock);
325
326         ENTRY;
327         spin_lock(&ns->ns_lock);
328         ldlm_lock_add_to_lru_nolock(lock);
329         spin_unlock(&ns->ns_lock);
330         EXIT;
331 }
332
333 /**
334  * Moves LDLM lock \a lock that is already in namespace LRU to the tail of
335  * the LRU. Performs necessary LRU locking
336  */
337 void ldlm_lock_touch_in_lru(struct ldlm_lock *lock)
338 {
339         struct ldlm_namespace *ns = ldlm_lock_to_ns(lock);
340
341         ENTRY;
342         if (ldlm_is_ns_srv(lock)) {
343                 LASSERT(list_empty(&lock->l_lru));
344                 EXIT;
345                 return;
346         }
347
348         spin_lock(&ns->ns_lock);
349         if (!list_empty(&lock->l_lru)) {
350                 ldlm_lock_remove_from_lru_nolock(lock);
351                 ldlm_lock_add_to_lru_nolock(lock);
352         }
353         spin_unlock(&ns->ns_lock);
354         EXIT;
355 }
356
357 /**
358  * Helper to destroy a locked lock.
359  *
360  * Used by ldlm_lock_destroy and ldlm_lock_destroy_nolock
361  * Must be called with l_lock and lr_lock held.
362  *
363  * Does not actually free the lock data, but rather marks the lock as
364  * destroyed by setting l_destroyed field in the lock to 1.  Destroys a
365  * handle->lock association too, so that the lock can no longer be found
366  * and removes the lock from LRU list.  Actual lock freeing occurs when
367  * last lock reference goes away.
368  *
369  * Original comment (of some historical value):
370  * This used to have a 'strict' flag, which recovery would use to mark an
371  * in-use lock as needing-to-die.  Lest I am ever tempted to put it back, I
372  * shall explain why it's gone: with the new hash table scheme, once you call
373  * ldlm_lock_destroy, you can never drop your final references on this lock.
374  * Because it's not in the hash table anymore.  -phil
375  */
376 static int ldlm_lock_destroy_internal(struct ldlm_lock *lock)
377 {
378         ENTRY;
379
380         if (lock->l_readers || lock->l_writers) {
381                 LDLM_ERROR(lock, "lock still has references");
382                 LBUG();
383         }
384
385         if (!list_empty(&lock->l_res_link)) {
386                 LDLM_ERROR(lock, "lock still on resource");
387                 LBUG();
388         }
389
390         if (ldlm_is_destroyed(lock)) {
391                 LASSERT(list_empty(&lock->l_lru));
392                 EXIT;
393                 return 0;
394         }
395         ldlm_set_destroyed(lock);
396
397         if (lock->l_export && lock->l_export->exp_lock_hash) {
398                 /* NB: it's safe to call cfs_hash_del() even lock isn't
399                  * in exp_lock_hash. */
400                 /* In the function below, .hs_keycmp resolves to
401                  * ldlm_export_lock_keycmp() */
402                 /* coverity[overrun-buffer-val] */
403                 cfs_hash_del(lock->l_export->exp_lock_hash,
404                              &lock->l_remote_handle, &lock->l_exp_hash);
405         }
406
407         ldlm_lock_remove_from_lru(lock);
408         class_handle_unhash(&lock->l_handle);
409
410         EXIT;
411         return 1;
412 }
413
414 /**
415  * Destroys a LDLM lock \a lock. Performs necessary locking first.
416  */
417 void ldlm_lock_destroy(struct ldlm_lock *lock)
418 {
419         int first;
420         ENTRY;
421         lock_res_and_lock(lock);
422         first = ldlm_lock_destroy_internal(lock);
423         unlock_res_and_lock(lock);
424
425         /* drop reference from hashtable only for first destroy */
426         if (first) {
427                 lu_ref_del(&lock->l_reference, "hash", lock);
428                 LDLM_LOCK_RELEASE(lock);
429         }
430         EXIT;
431 }
432
433 /**
434  * Destroys a LDLM lock \a lock that is already locked.
435  */
436 void ldlm_lock_destroy_nolock(struct ldlm_lock *lock)
437 {
438         int first;
439         ENTRY;
440         first = ldlm_lock_destroy_internal(lock);
441         /* drop reference from hashtable only for first destroy */
442         if (first) {
443                 lu_ref_del(&lock->l_reference, "hash", lock);
444                 LDLM_LOCK_RELEASE(lock);
445         }
446         EXIT;
447 }
448
449 static const char lock_handle_owner[] = "ldlm";
450
451 /**
452  *
453  * Allocate and initialize new lock structure.
454  *
455  * usage: pass in a resource on which you have done ldlm_resource_get
456  *        new lock will take over the refcount.
457  * returns: lock with refcount 2 - one for current caller and one for remote
458  */
459 static struct ldlm_lock *ldlm_lock_new(struct ldlm_resource *resource)
460 {
461         struct ldlm_lock *lock;
462         ENTRY;
463
464         if (resource == NULL)
465                 LBUG();
466
467         OBD_SLAB_ALLOC_PTR_GFP(lock, ldlm_lock_slab, GFP_NOFS);
468         if (lock == NULL)
469                 RETURN(NULL);
470
471         RCU_INIT_POINTER(lock->l_resource, resource);
472         lu_ref_add(&resource->lr_reference, "lock", lock);
473
474         refcount_set(&lock->l_handle.h_ref, 2);
475         INIT_LIST_HEAD(&lock->l_res_link);
476         INIT_LIST_HEAD(&lock->l_lru);
477         INIT_LIST_HEAD(&lock->l_pending_chain);
478         INIT_LIST_HEAD(&lock->l_bl_ast);
479         INIT_LIST_HEAD(&lock->l_cp_ast);
480         INIT_LIST_HEAD(&lock->l_rk_ast);
481         init_waitqueue_head(&lock->l_waitq);
482         lock->l_blocking_lock = NULL;
483         INIT_LIST_HEAD(&lock->l_sl_mode);
484         INIT_LIST_HEAD(&lock->l_sl_policy);
485         INIT_HLIST_NODE(&lock->l_exp_hash);
486         INIT_HLIST_NODE(&lock->l_exp_flock_hash);
487
488         lprocfs_counter_incr(ldlm_res_to_ns(resource)->ns_stats,
489                              LDLM_NSS_LOCKS);
490         INIT_HLIST_NODE(&lock->l_handle.h_link);
491         class_handle_hash(&lock->l_handle, lock_handle_owner);
492
493         lu_ref_init(&lock->l_reference);
494         lu_ref_add(&lock->l_reference, "hash", lock);
495         lock->l_callback_timestamp = 0;
496         lock->l_activity = 0;
497
498 #if LUSTRE_TRACKS_LOCK_EXP_REFS
499         INIT_LIST_HEAD(&lock->l_exp_refs_link);
500         lock->l_exp_refs_nr = 0;
501         lock->l_exp_refs_target = NULL;
502 #endif
503         INIT_LIST_HEAD(&lock->l_exp_list);
504
505         RETURN(lock);
506 }
507
508 /**
509  * Moves LDLM lock \a lock to another resource.
510  * This is used on client when server returns some other lock than requested
511  * (typically as a result of intent operation)
512  */
513 int ldlm_lock_change_resource(struct ldlm_namespace *ns, struct ldlm_lock *lock,
514                               const struct ldlm_res_id *new_resid)
515 {
516         struct ldlm_resource *oldres;
517         struct ldlm_resource *newres;
518         int type;
519         ENTRY;
520
521         LASSERT(ns_is_client(ns));
522
523         oldres = lock_res_and_lock(lock);
524         if (memcmp(new_resid, &oldres->lr_name,
525                    sizeof(oldres->lr_name)) == 0) {
526                 /* Nothing to do */
527                 unlock_res_and_lock(lock);
528                 RETURN(0);
529         }
530
531         LASSERT(new_resid->name[0] != 0);
532
533         /* This function assumes that the lock isn't on any lists */
534         LASSERT(list_empty(&lock->l_res_link));
535
536         type = oldres->lr_type;
537         unlock_res_and_lock(lock);
538
539         newres = ldlm_resource_get(ns, NULL, new_resid, type, 1);
540         if (IS_ERR(newres))
541                 RETURN(PTR_ERR(newres));
542
543         lu_ref_add(&newres->lr_reference, "lock", lock);
544         /*
545          * To flip the lock from the old to the new resource, oldres
546          * and newres have to be locked. Resource spin-locks are taken
547          * in the memory address order to avoid dead-locks.
548          * As this is the only circumstance where ->l_resource
549          * can change, and this cannot race with itself, it is safe
550          * to access lock->l_resource without being careful about locking.
551          */
552         oldres = lock->l_resource;
553         if (oldres < newres) {
554                 lock_res(oldres);
555                 lock_res_nested(newres, LRT_NEW);
556         } else {
557                 lock_res(newres);
558                 lock_res_nested(oldres, LRT_NEW);
559         }
560         LASSERT(memcmp(new_resid, &oldres->lr_name,
561                        sizeof oldres->lr_name) != 0);
562         rcu_assign_pointer(lock->l_resource, newres);
563         unlock_res(oldres);
564         unlock_res(newres);
565
566         /* ...and the flowers are still standing! */
567         lu_ref_del(&oldres->lr_reference, "lock", lock);
568         ldlm_resource_putref(oldres);
569
570         RETURN(0);
571 }
572
573 /** \defgroup ldlm_handles LDLM HANDLES
574  * Ways to get hold of locks without any addresses.
575  * @{
576  */
577
578 /**
579  * Fills in handle for LDLM lock \a lock into supplied \a lockh
580  * Does not take any references.
581  */
582 void ldlm_lock2handle(const struct ldlm_lock *lock, struct lustre_handle *lockh)
583 {
584         lockh->cookie = lock->l_handle.h_cookie;
585 }
586 EXPORT_SYMBOL(ldlm_lock2handle);
587
588 /**
589  * Obtain a lock reference by handle.
590  *
591  * if \a flags: atomically get the lock and set the flags.
592  *              Return NULL if flag already set
593  */
594 struct ldlm_lock *__ldlm_handle2lock(const struct lustre_handle *handle,
595                                      __u64 flags)
596 {
597         struct ldlm_lock *lock;
598         ENTRY;
599
600         LASSERT(handle);
601
602         if (!lustre_handle_is_used(handle))
603                 RETURN(NULL);
604
605         lock = class_handle2object(handle->cookie, lock_handle_owner);
606         if (lock == NULL)
607                 RETURN(NULL);
608
609         if (lock->l_export != NULL && lock->l_export->exp_failed) {
610                 CDEBUG(D_INFO, "lock export failed: lock %p, exp %p\n",
611                        lock, lock->l_export);
612                 LDLM_LOCK_PUT(lock);
613                 RETURN(NULL);
614         }
615
616         /* It's unlikely but possible that someone marked the lock as
617          * destroyed after we did handle2object on it */
618         if ((flags == 0) && !ldlm_is_destroyed(lock)) {
619                 lu_ref_add_atomic(&lock->l_reference, "handle", lock);
620                 RETURN(lock);
621         }
622
623         lock_res_and_lock(lock);
624
625         LASSERT(lock->l_resource != NULL);
626
627         lu_ref_add_atomic(&lock->l_reference, "handle", lock);
628         if (unlikely(ldlm_is_destroyed(lock))) {
629                 unlock_res_and_lock(lock);
630                 CDEBUG(D_INFO, "lock already destroyed: lock %p\n", lock);
631                 LDLM_LOCK_PUT(lock);
632                 RETURN(NULL);
633         }
634
635         /* If we're setting flags, make sure none of them are already set. */
636         if (flags != 0) {
637                 if ((lock->l_flags & flags) != 0) {
638                         unlock_res_and_lock(lock);
639                         LDLM_LOCK_PUT(lock);
640                         RETURN(NULL);
641                 }
642
643                 lock->l_flags |= flags;
644         }
645
646         unlock_res_and_lock(lock);
647         RETURN(lock);
648 }
649 EXPORT_SYMBOL(__ldlm_handle2lock);
650 /** @} ldlm_handles */
651
652 /**
653  * Fill in "on the wire" representation for given LDLM lock into supplied
654  * lock descriptor \a desc structure.
655  */
656 void ldlm_lock2desc(struct ldlm_lock *lock, struct ldlm_lock_desc *desc)
657 {
658         ldlm_res2desc(lock->l_resource, &desc->l_resource);
659         desc->l_req_mode = lock->l_req_mode;
660         desc->l_granted_mode = lock->l_granted_mode;
661         ldlm_convert_policy_to_wire(lock->l_resource->lr_type,
662                                     &lock->l_policy_data,
663                                     &desc->l_policy_data);
664 }
665
666 /**
667  * Add a lock to list of conflicting locks to send AST to.
668  *
669  * Only add if we have not sent a blocking AST to the lock yet.
670  */
671 static void ldlm_add_bl_work_item(struct ldlm_lock *lock, struct ldlm_lock *new,
672                                   struct list_head *work_list)
673 {
674         if (!ldlm_is_ast_sent(lock)) {
675                 LDLM_DEBUG(lock, "lock incompatible; sending blocking AST.");
676                 ldlm_set_ast_sent(lock);
677                 /* If the enqueuing client said so, tell the AST recipient to
678                  * discard dirty data, rather than writing back. */
679                 if (ldlm_is_ast_discard_data(new))
680                         ldlm_set_discard_data(lock);
681
682                 /* Lock can be converted from a blocking state back to granted
683                  * after lock convert or COS downgrade but still be in an
684                  * older bl_list because it is controlled only by
685                  * ldlm_work_bl_ast_lock(), let it be processed there.
686                  */
687                 if (list_empty(&lock->l_bl_ast)) {
688                         list_add(&lock->l_bl_ast, work_list);
689                         LDLM_LOCK_GET(lock);
690                 }
691                 LASSERT(lock->l_blocking_lock == NULL);
692                 lock->l_blocking_lock = LDLM_LOCK_GET(new);
693         }
694 }
695
696 /**
697  * Add a lock to list of just granted locks to send completion AST to.
698  */
699 static void ldlm_add_cp_work_item(struct ldlm_lock *lock,
700                                   struct list_head *work_list)
701 {
702         if (!ldlm_is_cp_reqd(lock)) {
703                 ldlm_set_cp_reqd(lock);
704                 LDLM_DEBUG(lock, "lock granted; sending completion AST.");
705                 LASSERT(list_empty(&lock->l_cp_ast));
706                 list_add(&lock->l_cp_ast, work_list);
707                 LDLM_LOCK_GET(lock);
708         }
709 }
710
711 /**
712  * Aggregator function to add AST work items into a list. Determines
713  * what sort of an AST work needs to be done and calls the proper
714  * adding function.
715  * Must be called with lr_lock held.
716  */
717 void ldlm_add_ast_work_item(struct ldlm_lock *lock, struct ldlm_lock *new,
718                             struct list_head *work_list)
719 {
720         ENTRY;
721         check_res_locked(lock->l_resource);
722         if (new)
723                 ldlm_add_bl_work_item(lock, new, work_list);
724         else
725                 ldlm_add_cp_work_item(lock, work_list);
726         EXIT;
727 }
728
729 /**
730  * Add specified reader/writer reference to LDLM lock with handle \a lockh.
731  * r/w reference type is determined by \a mode
732  * Calls ldlm_lock_addref_internal.
733  */
734 void ldlm_lock_addref(const struct lustre_handle *lockh, enum ldlm_mode mode)
735 {
736         struct ldlm_lock *lock;
737
738         lock = ldlm_handle2lock(lockh);
739         LASSERTF(lock != NULL, "Non-existing lock: %#llx\n", lockh->cookie);
740         ldlm_lock_addref_internal(lock, mode);
741         LDLM_LOCK_PUT(lock);
742 }
743 EXPORT_SYMBOL(ldlm_lock_addref);
744
745 /**
746  * Helper function.
747  * Add specified reader/writer reference to LDLM lock \a lock.
748  * r/w reference type is determined by \a mode
749  * Removes lock from LRU if it is there.
750  * Assumes the LDLM lock is already locked.
751  */
752 void ldlm_lock_addref_internal_nolock(struct ldlm_lock *lock,
753                                       enum ldlm_mode mode)
754 {
755         ldlm_lock_remove_from_lru(lock);
756         if (mode & (LCK_NL | LCK_CR | LCK_PR)) {
757                 lock->l_readers++;
758                 lu_ref_add_atomic(&lock->l_reference, "reader", lock);
759         }
760         if (mode & (LCK_EX | LCK_CW | LCK_PW | LCK_GROUP | LCK_COS)) {
761                 lock->l_writers++;
762                 lu_ref_add_atomic(&lock->l_reference, "writer", lock);
763         }
764         LDLM_LOCK_GET(lock);
765         lu_ref_add_atomic(&lock->l_reference, "user", lock);
766         LDLM_DEBUG(lock, "ldlm_lock_addref(%s)", ldlm_lockname[mode]);
767 }
768
769 /**
770  * Attempts to add reader/writer reference to a lock with handle \a lockh, and
771  * fails if lock is already LDLM_FL_CBPENDING or destroyed.
772  *
773  * \retval 0 success, lock was addref-ed
774  *
775  * \retval -EAGAIN lock is being canceled.
776  */
777 int ldlm_lock_addref_try(const struct lustre_handle *lockh, enum ldlm_mode mode)
778 {
779         struct ldlm_lock *lock;
780         int               result;
781
782         result = -EAGAIN;
783         lock = ldlm_handle2lock(lockh);
784         if (lock != NULL) {
785                 lock_res_and_lock(lock);
786                 if (lock->l_readers != 0 || lock->l_writers != 0 ||
787                     !ldlm_is_cbpending(lock)) {
788                         ldlm_lock_addref_internal_nolock(lock, mode);
789                         result = 0;
790                 }
791                 unlock_res_and_lock(lock);
792                 LDLM_LOCK_PUT(lock);
793         }
794         return result;
795 }
796 EXPORT_SYMBOL(ldlm_lock_addref_try);
797
798 /**
799  * Add specified reader/writer reference to LDLM lock \a lock.
800  * Locks LDLM lock and calls ldlm_lock_addref_internal_nolock to do the work.
801  * Only called for local locks.
802  */
803 void ldlm_lock_addref_internal(struct ldlm_lock *lock, enum ldlm_mode mode)
804 {
805         lock_res_and_lock(lock);
806         ldlm_lock_addref_internal_nolock(lock, mode);
807         unlock_res_and_lock(lock);
808 }
809
810 /**
811  * Removes reader/writer reference for LDLM lock \a lock.
812  * Assumes LDLM lock is already locked.
813  * only called in ldlm_flock_destroy and for local locks.
814  * Does NOT add lock to LRU if no r/w references left to accomodate flock locks
815  * that cannot be placed in LRU.
816  */
817 void ldlm_lock_decref_internal_nolock(struct ldlm_lock *lock,
818                                       enum ldlm_mode mode)
819 {
820         LDLM_DEBUG(lock, "ldlm_lock_decref(%s)", ldlm_lockname[mode]);
821         if (mode & (LCK_NL | LCK_CR | LCK_PR)) {
822                 LASSERT(lock->l_readers > 0);
823                 lu_ref_del(&lock->l_reference, "reader", lock);
824                 lock->l_readers--;
825         }
826         if (mode & (LCK_EX | LCK_CW | LCK_PW | LCK_GROUP | LCK_COS)) {
827                 LASSERT(lock->l_writers > 0);
828                 lu_ref_del(&lock->l_reference, "writer", lock);
829                 lock->l_writers--;
830         }
831
832         lu_ref_del(&lock->l_reference, "user", lock);
833         LDLM_LOCK_RELEASE(lock);    /* matches the LDLM_LOCK_GET() in addref */
834 }
835
836 /**
837  * Removes reader/writer reference for LDLM lock \a lock.
838  * Locks LDLM lock first.
839  * If the lock is determined to be client lock on a client and r/w refcount
840  * drops to zero and the lock is not blocked, the lock is added to LRU lock
841  * on the namespace.
842  * For blocked LDLM locks if r/w count drops to zero, blocking_ast is called.
843  */
844 void ldlm_lock_decref_internal(struct ldlm_lock *lock, enum ldlm_mode mode)
845 {
846         struct ldlm_namespace *ns;
847
848         ENTRY;
849
850         lock_res_and_lock(lock);
851
852         ns = ldlm_lock_to_ns(lock);
853
854         ldlm_lock_decref_internal_nolock(lock, mode);
855
856         if ((ldlm_is_local(lock) || lock->l_req_mode == LCK_GROUP) &&
857             !lock->l_readers && !lock->l_writers) {
858                 /* If this is a local lock on a server namespace and this was
859                  * the last reference, cancel the lock.
860                  *
861                  * Group locks are special:
862                  * They must not go in LRU, but they are not called back
863                  * like non-group locks, instead they are manually released.
864                  * They have an l_writers reference which they keep until
865                  * they are manually released, so we remove them when they have
866                  * no more reader or writer references. - LU-6368 */
867                 ldlm_set_cbpending(lock);
868         }
869
870         if (!lock->l_readers && !lock->l_writers && ldlm_is_cbpending(lock)) {
871                 unsigned int mask = D_DLMTRACE;
872
873                 /* If we received a blocked AST and this was the last reference,
874                  * run the callback. */
875                 if (ldlm_is_ns_srv(lock) && lock->l_export)
876                         mask |= D_WARNING;
877                 LDLM_DEBUG_LIMIT(mask, lock,
878                                  "final decref done on %sCBPENDING lock",
879                                  mask & D_WARNING ? "non-local " : "");
880
881                 LDLM_LOCK_GET(lock); /* dropped by bl thread */
882                 ldlm_lock_remove_from_lru(lock);
883                 unlock_res_and_lock(lock);
884
885                 if (ldlm_is_fail_loc(lock))
886                         OBD_RACE(OBD_FAIL_LDLM_CP_BL_RACE);
887
888                 if (ldlm_is_atomic_cb(lock) ||
889                     ldlm_bl_to_thread_lock(ns, NULL, lock) != 0)
890                         ldlm_handle_bl_callback(ns, NULL, lock);
891         } else if (ns_is_client(ns) &&
892                    !lock->l_readers && !lock->l_writers &&
893                    !ldlm_is_no_lru(lock) &&
894                    !ldlm_is_bl_ast(lock) &&
895                    !ldlm_is_converting(lock)) {
896
897                 /* If this is a client-side namespace and this was the last
898                  * reference, put it on the LRU.
899                  */
900                 ldlm_lock_add_to_lru(lock);
901                 unlock_res_and_lock(lock);
902                 LDLM_DEBUG(lock, "add lock into lru list");
903
904                 if (ldlm_is_fail_loc(lock))
905                         OBD_RACE(OBD_FAIL_LDLM_CP_BL_RACE);
906
907                 ldlm_pool_recalc(&ns->ns_pool, true);
908         } else {
909                 LDLM_DEBUG(lock, "do not add lock into lru list");
910                 unlock_res_and_lock(lock);
911         }
912
913         EXIT;
914 }
915
916 /**
917  * Decrease reader/writer refcount for LDLM lock with handle \a lockh
918  */
919 void ldlm_lock_decref(const struct lustre_handle *lockh, enum ldlm_mode mode)
920 {
921         struct ldlm_lock *lock = __ldlm_handle2lock(lockh, 0);
922         LASSERTF(lock != NULL, "Non-existing lock: %#llx\n", lockh->cookie);
923         ldlm_lock_decref_internal(lock, mode);
924         LDLM_LOCK_PUT(lock);
925 }
926 EXPORT_SYMBOL(ldlm_lock_decref);
927
928 /**
929  * Decrease reader/writer refcount for LDLM lock with handle
930  * \a lockh and mark it for subsequent cancellation once r/w refcount
931  * drops to zero instead of putting into LRU.
932  *
933  */
934 void ldlm_lock_decref_and_cancel(const struct lustre_handle *lockh,
935                                  enum ldlm_mode mode)
936 {
937         struct ldlm_lock *lock = __ldlm_handle2lock(lockh, 0);
938         ENTRY;
939
940         LASSERT(lock != NULL);
941
942         LDLM_DEBUG(lock, "ldlm_lock_decref(%s)", ldlm_lockname[mode]);
943         lock_res_and_lock(lock);
944         ldlm_set_cbpending(lock);
945         unlock_res_and_lock(lock);
946         ldlm_lock_decref_internal(lock, mode);
947         LDLM_LOCK_PUT(lock);
948 }
949 EXPORT_SYMBOL(ldlm_lock_decref_and_cancel);
950
951 struct sl_insert_point {
952         struct list_head *res_link;
953         struct list_head *mode_link;
954         struct list_head *policy_link;
955 };
956
957 /**
958  * Finds a position to insert the new lock into granted lock list.
959  *
960  * Used for locks eligible for skiplist optimization.
961  *
962  * Parameters:
963  *      queue [input]:  the granted list where search acts on;
964  *      req [input]:    the lock whose position to be located;
965  *      prev [output]:  positions within 3 lists to insert @req to
966  * Return Value:
967  *      filled @prev
968  * NOTE: called by
969  *  - ldlm_grant_lock_with_skiplist
970  */
971 static void search_granted_lock(struct list_head *queue,
972                                 struct ldlm_lock *req,
973                                 struct sl_insert_point *prev)
974 {
975         struct list_head *tmp;
976         struct ldlm_lock *lock, *mode_end, *policy_end;
977         ENTRY;
978
979         list_for_each(tmp, queue) {
980                 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
981
982                 mode_end = list_entry(lock->l_sl_mode.prev,
983                                           struct ldlm_lock, l_sl_mode);
984
985                 if (lock->l_req_mode != req->l_req_mode) {
986                         /* jump to last lock of mode group */
987                         tmp = &mode_end->l_res_link;
988                         continue;
989                 }
990
991                 /* suitable mode group is found */
992                 if (lock->l_resource->lr_type == LDLM_PLAIN) {
993                         /* insert point is last lock of the mode group */
994                         prev->res_link = &mode_end->l_res_link;
995                         prev->mode_link = &mode_end->l_sl_mode;
996                         prev->policy_link = &req->l_sl_policy;
997                         EXIT;
998                         return;
999                 } else if (lock->l_resource->lr_type == LDLM_IBITS) {
1000                         for (;;) {
1001                                 policy_end =
1002                                         list_entry(lock->l_sl_policy.prev,
1003                                                        struct ldlm_lock,
1004                                                        l_sl_policy);
1005
1006                                 if (lock->l_policy_data.l_inodebits.bits ==
1007                                     req->l_policy_data.l_inodebits.bits) {
1008                                         /* insert point is last lock of
1009                                          * the policy group */
1010                                         prev->res_link =
1011                                                 &policy_end->l_res_link;
1012                                         prev->mode_link =
1013                                                 &policy_end->l_sl_mode;
1014                                         prev->policy_link =
1015                                                 &policy_end->l_sl_policy;
1016                                         EXIT;
1017                                         return;
1018                                 }
1019
1020                                 if (policy_end == mode_end)
1021                                         /* done with mode group */
1022                                         break;
1023
1024                                 /* go to next policy group within mode group */
1025                                 tmp = policy_end->l_res_link.next;
1026                                 lock = list_entry(tmp, struct ldlm_lock,
1027                                                       l_res_link);
1028                         }  /* loop over policy groups within the mode group */
1029
1030                         /* insert point is last lock of the mode group,
1031                          * new policy group is started */
1032                         prev->res_link = &mode_end->l_res_link;
1033                         prev->mode_link = &mode_end->l_sl_mode;
1034                         prev->policy_link = &req->l_sl_policy;
1035                         EXIT;
1036                         return;
1037                 } else {
1038                         LDLM_ERROR(lock,"is not LDLM_PLAIN or LDLM_IBITS lock");
1039                         LBUG();
1040                 }
1041         }
1042
1043         /* insert point is last lock on the queue,
1044          * new mode group and new policy group are started */
1045         prev->res_link = queue->prev;
1046         prev->mode_link = &req->l_sl_mode;
1047         prev->policy_link = &req->l_sl_policy;
1048         EXIT;
1049 }
1050
1051 /**
1052  * Add a lock into resource granted list after a position described by
1053  * \a prev.
1054  */
1055 static void ldlm_granted_list_add_lock(struct ldlm_lock *lock,
1056                                        struct sl_insert_point *prev)
1057 {
1058         struct ldlm_resource *res = lock->l_resource;
1059         ENTRY;
1060
1061         check_res_locked(res);
1062
1063         ldlm_resource_dump(D_INFO, res);
1064         LDLM_DEBUG(lock, "About to add lock:");
1065
1066         if (ldlm_is_destroyed(lock)) {
1067                 CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n");
1068                 return;
1069         }
1070
1071         LASSERT(list_empty(&lock->l_res_link));
1072         LASSERT(list_empty(&lock->l_sl_mode));
1073         LASSERT(list_empty(&lock->l_sl_policy));
1074
1075         /*
1076          * lock->link == prev->link means lock is first starting the group.
1077          * Don't re-add to itself to suppress kernel warnings.
1078          */
1079         if (&lock->l_res_link != prev->res_link)
1080                 list_add(&lock->l_res_link, prev->res_link);
1081         if (&lock->l_sl_mode != prev->mode_link)
1082                 list_add(&lock->l_sl_mode, prev->mode_link);
1083         if (&lock->l_sl_policy != prev->policy_link)
1084                 list_add(&lock->l_sl_policy, prev->policy_link);
1085
1086         EXIT;
1087 }
1088
1089 /**
1090  * Add a lock to granted list on a resource maintaining skiplist
1091  * correctness.
1092  */
1093 void ldlm_grant_lock_with_skiplist(struct ldlm_lock *lock)
1094 {
1095         struct sl_insert_point prev;
1096
1097         LASSERT(ldlm_is_granted(lock));
1098
1099         search_granted_lock(&lock->l_resource->lr_granted, lock, &prev);
1100         ldlm_granted_list_add_lock(lock, &prev);
1101 }
1102
1103 /**
1104  * Perform lock granting bookkeeping.
1105  *
1106  * Includes putting the lock into granted list and updating lock mode.
1107  * NOTE: called by
1108  *  - ldlm_lock_enqueue
1109  *  - ldlm_reprocess_queue
1110  *
1111  * must be called with lr_lock held
1112  */
1113 void ldlm_grant_lock(struct ldlm_lock *lock, struct list_head *work_list)
1114 {
1115         struct ldlm_resource *res = lock->l_resource;
1116         ENTRY;
1117
1118         check_res_locked(res);
1119
1120         lock->l_granted_mode = lock->l_req_mode;
1121
1122         if (work_list && lock->l_completion_ast != NULL)
1123                 ldlm_add_ast_work_item(lock, NULL, work_list);
1124
1125         if (res->lr_type == LDLM_PLAIN || res->lr_type == LDLM_IBITS)
1126                 ldlm_grant_lock_with_skiplist(lock);
1127         else if (res->lr_type == LDLM_EXTENT)
1128                 ldlm_extent_add_lock(res, lock);
1129         else if (res->lr_type == LDLM_FLOCK) {
1130                 /* We should not add locks to granted list in the following
1131                  * cases:
1132                  * - this is an UNLOCK but not a real lock;
1133                  * - this is a TEST lock;
1134                  * - this is a F_CANCELLK lock (async flock has req_mode == 0)
1135                  * - this is a deadlock (flock cannot be granted) */
1136                 if (lock->l_req_mode == 0 ||
1137                     lock->l_req_mode == LCK_NL ||
1138                     ldlm_is_test_lock(lock) ||
1139                     ldlm_is_flock_deadlock(lock))
1140                         RETURN_EXIT;
1141                 ldlm_resource_add_lock(res, &res->lr_granted, lock);
1142         } else {
1143                 LBUG();
1144         }
1145
1146         ldlm_pool_add(&ldlm_res_to_ns(res)->ns_pool, lock);
1147         EXIT;
1148 }
1149
1150 /**
1151  * Check if the given @lock meets the criteria for a match.
1152  * A reference on the lock is taken if matched.
1153  *
1154  * @lock        test-against this lock
1155  * @data        parameters
1156  *
1157  * RETURN       returns true if @lock matches @data, false otherwise
1158  */
1159 static bool lock_matches(struct ldlm_lock *lock, struct ldlm_match_data *data)
1160 {
1161         union ldlm_policy_data *lpol = &lock->l_policy_data;
1162         enum ldlm_mode match = LCK_MINMODE;
1163
1164         if (lock == data->lmd_old)
1165                 return true;
1166
1167         /* Check if this lock can be matched.
1168          * Used by LU-2919(exclusive open) for open lease lock */
1169         if (ldlm_is_excl(lock))
1170                 return false;
1171
1172         /* llite sometimes wants to match locks that will be
1173          * canceled when their users drop, but we allow it to match
1174          * if it passes in CBPENDING and the lock still has users.
1175          * this is generally only going to be used by children
1176          * whose parents already hold a lock so forward progress
1177          * can still happen. */
1178         if (ldlm_is_cbpending(lock) &&
1179             !(data->lmd_flags & LDLM_FL_CBPENDING))
1180                 return false;
1181
1182         if (!(data->lmd_match & LDLM_MATCH_UNREF) && ldlm_is_cbpending(lock) &&
1183             lock->l_readers == 0 && lock->l_writers == 0)
1184                 return false;
1185
1186         if (!(lock->l_req_mode & *data->lmd_mode))
1187                 return false;
1188
1189         /* When we search for ast_data, we are not doing a traditional match,
1190          * so we don't worry about IBITS or extent matching.
1191          */
1192         if (data->lmd_match & (LDLM_MATCH_AST | LDLM_MATCH_AST_ANY)) {
1193                 if (!lock->l_ast_data)
1194                         return false;
1195
1196                 if (data->lmd_match & LDLM_MATCH_AST_ANY)
1197                         goto matched;
1198         }
1199
1200         match = lock->l_req_mode;
1201
1202         switch (lock->l_resource->lr_type) {
1203         case LDLM_EXTENT:
1204                 if (!(data->lmd_match & LDLM_MATCH_RIGHT) &&
1205                     (lpol->l_extent.start > data->lmd_policy->l_extent.start ||
1206                      lpol->l_extent.end < data->lmd_policy->l_extent.end))
1207                         return false;
1208
1209                 if (unlikely(match == LCK_GROUP) &&
1210                     data->lmd_policy->l_extent.gid != LDLM_GID_ANY &&
1211                     lpol->l_extent.gid != data->lmd_policy->l_extent.gid)
1212                         return false;
1213                 break;
1214         case LDLM_IBITS:
1215                 /* We match if we have existing lock with same or wider set
1216                    of bits. */
1217                 if ((lpol->l_inodebits.bits &
1218                      data->lmd_policy->l_inodebits.bits) !=
1219                     data->lmd_policy->l_inodebits.bits)
1220                         return false;
1221
1222                 if (unlikely(match == LCK_GROUP) &&
1223                     data->lmd_policy->l_inodebits.li_gid != LDLM_GID_ANY &&
1224                     lpol->l_inodebits.li_gid !=
1225                     data->lmd_policy->l_inodebits.li_gid)
1226                         return false;
1227                 break;
1228         default:
1229                 ;
1230         }
1231
1232         /* We match if we have existing lock with same or wider set
1233            of bits. */
1234         if (!(data->lmd_match & LDLM_MATCH_UNREF) && LDLM_HAVE_MASK(lock, GONE))
1235                 return false;
1236
1237         if (!equi(data->lmd_flags & LDLM_FL_LOCAL_ONLY, ldlm_is_local(lock)))
1238                 return false;
1239
1240         /* Filter locks by skipping flags */
1241         if (data->lmd_skip_flags & lock->l_flags)
1242                 return false;
1243
1244 matched:
1245         if (data->lmd_flags & LDLM_FL_TEST_LOCK) {
1246                 LDLM_LOCK_GET(lock);
1247                 ldlm_lock_touch_in_lru(lock);
1248         } else {
1249                 ldlm_lock_addref_internal_nolock(lock, match);
1250         }
1251
1252         *data->lmd_mode = match;
1253         data->lmd_lock = lock;
1254
1255         return true;
1256 }
1257
1258 static unsigned int itree_overlap_cb(struct interval_node *in, void *args)
1259 {
1260         struct ldlm_interval *node = to_ldlm_interval(in);
1261         struct ldlm_match_data *data = args;
1262         struct ldlm_lock *lock;
1263
1264         list_for_each_entry(lock, &node->li_group, l_sl_policy) {
1265                 if (lock_matches(lock, data))
1266                         return INTERVAL_ITER_STOP;
1267         }
1268         return INTERVAL_ITER_CONT;
1269 }
1270
1271 /**
1272  * Search for a lock with given parameters in interval trees.
1273  *
1274  * \param res      search for a lock in this resource
1275  * \param data     parameters
1276  *
1277  * \retval a referenced lock or NULL.
1278  */
1279 struct ldlm_lock *search_itree(struct ldlm_resource *res,
1280                                struct ldlm_match_data *data)
1281 {
1282         struct interval_node_extent ext = {
1283                 .start     = data->lmd_policy->l_extent.start,
1284                 .end       = data->lmd_policy->l_extent.end
1285         };
1286         int idx;
1287
1288         data->lmd_lock = NULL;
1289
1290         if (data->lmd_match & LDLM_MATCH_RIGHT)
1291                 ext.end = OBD_OBJECT_EOF;
1292
1293         for (idx = 0; idx < LCK_MODE_NUM; idx++) {
1294                 struct ldlm_interval_tree *tree = &res->lr_itree[idx];
1295
1296                 if (tree->lit_root == NULL)
1297                         continue;
1298
1299                 if (!(tree->lit_mode & *data->lmd_mode))
1300                         continue;
1301
1302                 interval_search(tree->lit_root, &ext,
1303                                 itree_overlap_cb, data);
1304                 if (data->lmd_lock)
1305                         return data->lmd_lock;
1306         }
1307
1308         return NULL;
1309 }
1310 EXPORT_SYMBOL(search_itree);
1311
1312
1313 /**
1314  * Search for a lock with given properties in a queue.
1315  *
1316  * \param queue    search for a lock in this queue
1317  * \param data     parameters
1318  *
1319  * \retval a referenced lock or NULL.
1320  */
1321 static struct ldlm_lock *search_queue(struct list_head *queue,
1322                                       struct ldlm_match_data *data)
1323 {
1324         struct ldlm_lock *lock;
1325
1326         data->lmd_lock = NULL;
1327
1328         list_for_each_entry(lock, queue, l_res_link)
1329                 if (lock_matches(lock, data))
1330                         return data->lmd_lock;
1331
1332         return NULL;
1333 }
1334
1335 void ldlm_lock_fail_match_locked(struct ldlm_lock *lock)
1336 {
1337         if ((lock->l_flags & LDLM_FL_FAIL_NOTIFIED) == 0) {
1338                 lock->l_flags |= LDLM_FL_FAIL_NOTIFIED;
1339                 wake_up(&lock->l_waitq);
1340         }
1341 }
1342 EXPORT_SYMBOL(ldlm_lock_fail_match_locked);
1343
1344 void ldlm_lock_fail_match(struct ldlm_lock *lock)
1345 {
1346         lock_res_and_lock(lock);
1347         ldlm_lock_fail_match_locked(lock);
1348         unlock_res_and_lock(lock);
1349 }
1350
1351 /**
1352  * Mark lock as "matchable" by OST.
1353  *
1354  * Used to prevent certain races in LOV/OSC where the lock is granted, but LVB
1355  * is not yet valid.
1356  * Assumes LDLM lock is already locked.
1357  */
1358 void ldlm_lock_allow_match_locked(struct ldlm_lock *lock)
1359 {
1360         ldlm_set_lvb_ready(lock);
1361         wake_up(&lock->l_waitq);
1362 }
1363 EXPORT_SYMBOL(ldlm_lock_allow_match_locked);
1364
1365 /**
1366  * Mark lock as "matchable" by OST.
1367  * Locks the lock and then \see ldlm_lock_allow_match_locked
1368  */
1369 void ldlm_lock_allow_match(struct ldlm_lock *lock)
1370 {
1371         lock_res_and_lock(lock);
1372         ldlm_lock_allow_match_locked(lock);
1373         unlock_res_and_lock(lock);
1374 }
1375 EXPORT_SYMBOL(ldlm_lock_allow_match);
1376
1377 /**
1378  * Attempt to find a lock with specified properties.
1379  *
1380  * Typically returns a reference to matched lock unless LDLM_FL_TEST_LOCK is
1381  * set in \a flags
1382  *
1383  * Can be called in two ways:
1384  *
1385  * If 'ns' is NULL, then lockh describes an existing lock that we want to look
1386  * for a duplicate of.
1387  *
1388  * Otherwise, all of the fields must be filled in, to match against.
1389  *
1390  * If 'flags' contains LDLM_FL_LOCAL_ONLY, then only match local locks on the
1391  *     server (ie, connh is NULL)
1392  * If 'flags' contains LDLM_FL_BLOCK_GRANTED, then only locks on the granted
1393  *     list will be considered
1394  * If 'flags' contains LDLM_FL_CBPENDING, then locks that have been marked
1395  *     to be canceled can still be matched as long as they still have reader
1396  *     or writer refernces
1397  * If 'flags' contains LDLM_FL_TEST_LOCK, then don't actually reference a lock,
1398  *     just tell us if we would have matched.
1399  *
1400  * \retval 1 if it finds an already-existing lock that is compatible; in this
1401  * case, lockh is filled in with a addref()ed lock
1402  *
1403  * We also check security context, and if that fails we simply return 0 (to
1404  * keep caller code unchanged), the context failure will be discovered by
1405  * caller sometime later.
1406  */
1407 enum ldlm_mode ldlm_lock_match_with_skip(struct ldlm_namespace *ns,
1408                                          __u64 flags, __u64 skip_flags,
1409                                          const struct ldlm_res_id *res_id,
1410                                          enum ldlm_type type,
1411                                          union ldlm_policy_data *policy,
1412                                          enum ldlm_mode mode,
1413                                          struct lustre_handle *lockh,
1414                                          enum ldlm_match_flags match_flags)
1415 {
1416         struct ldlm_match_data data = {
1417                 .lmd_old = NULL,
1418                 .lmd_lock = NULL,
1419                 .lmd_mode = &mode,
1420                 .lmd_policy = policy,
1421                 .lmd_flags = flags,
1422                 .lmd_skip_flags = skip_flags,
1423                 .lmd_match = match_flags,
1424         };
1425         struct ldlm_resource *res;
1426         struct ldlm_lock *lock;
1427         int matched;
1428
1429         ENTRY;
1430
1431         if (ns == NULL) {
1432                 data.lmd_old = ldlm_handle2lock(lockh);
1433                 LASSERT(data.lmd_old != NULL);
1434
1435                 ns = ldlm_lock_to_ns(data.lmd_old);
1436                 res_id = &data.lmd_old->l_resource->lr_name;
1437                 type = data.lmd_old->l_resource->lr_type;
1438                 *data.lmd_mode = data.lmd_old->l_req_mode;
1439         }
1440
1441         res = ldlm_resource_get(ns, NULL, res_id, type, 0);
1442         if (IS_ERR(res)) {
1443                 LASSERT(data.lmd_old == NULL);
1444                 RETURN(0);
1445         }
1446
1447         LDLM_RESOURCE_ADDREF(res);
1448         lock_res(res);
1449         if (res->lr_type == LDLM_EXTENT)
1450                 lock = search_itree(res, &data);
1451         else
1452                 lock = search_queue(&res->lr_granted, &data);
1453         if (!lock && !(flags & LDLM_FL_BLOCK_GRANTED))
1454                 lock = search_queue(&res->lr_waiting, &data);
1455         matched = lock ? mode : 0;
1456         unlock_res(res);
1457         LDLM_RESOURCE_DELREF(res);
1458         ldlm_resource_putref(res);
1459
1460         if (lock) {
1461                 ldlm_lock2handle(lock, lockh);
1462                 if ((flags & LDLM_FL_LVB_READY) &&
1463                     (!ldlm_is_lvb_ready(lock))) {
1464                         __u64 wait_flags = LDLM_FL_LVB_READY |
1465                                 LDLM_FL_DESTROYED | LDLM_FL_FAIL_NOTIFIED;
1466
1467                         if (lock->l_completion_ast) {
1468                                 int err = lock->l_completion_ast(lock,
1469                                                         LDLM_FL_WAIT_NOREPROC,
1470                                                         NULL);
1471                                 if (err)
1472                                         GOTO(out_fail_match, matched = 0);
1473                         }
1474
1475                         wait_event_idle_timeout(
1476                                 lock->l_waitq,
1477                                 lock->l_flags & wait_flags,
1478                                 cfs_time_seconds(obd_timeout));
1479
1480                         if (!ldlm_is_lvb_ready(lock))
1481                                 GOTO(out_fail_match, matched = 0);
1482                 }
1483
1484                 /* check user's security context */
1485                 if (lock->l_conn_export &&
1486                     sptlrpc_import_check_ctx(
1487                                 class_exp2cliimp(lock->l_conn_export)))
1488                         GOTO(out_fail_match, matched = 0);
1489
1490                 LDLM_DEBUG(lock, "matched (%llu %llu)",
1491                            (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1492                            res_id->name[2] : policy->l_extent.start,
1493                            (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1494                            res_id->name[3] : policy->l_extent.end);
1495
1496 out_fail_match:
1497                 if (flags & LDLM_FL_TEST_LOCK)
1498                         LDLM_LOCK_RELEASE(lock);
1499                 else if (!matched)
1500                         ldlm_lock_decref_internal(lock, mode);
1501         }
1502
1503         /* less verbose for test-only */
1504         if (!matched && !(flags & LDLM_FL_TEST_LOCK)) {
1505                 LDLM_DEBUG_NOLOCK("not matched ns %p type %u mode %u res "
1506                                   "%llu/%llu (%llu %llu)", ns,
1507                                   type, mode, res_id->name[0], res_id->name[1],
1508                                   (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1509                                   res_id->name[2] : policy->l_extent.start,
1510                                   (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1511                                   res_id->name[3] : policy->l_extent.end);
1512         }
1513         if (data.lmd_old != NULL)
1514                 LDLM_LOCK_PUT(data.lmd_old);
1515
1516         return matched;
1517 }
1518 EXPORT_SYMBOL(ldlm_lock_match_with_skip);
1519
1520 enum ldlm_mode ldlm_revalidate_lock_handle(const struct lustre_handle *lockh,
1521                                            __u64 *bits)
1522 {
1523         struct ldlm_lock *lock;
1524         enum ldlm_mode mode = 0;
1525         ENTRY;
1526
1527         lock = ldlm_handle2lock(lockh);
1528         if (lock != NULL) {
1529                 lock_res_and_lock(lock);
1530                 if (LDLM_HAVE_MASK(lock, GONE))
1531                         GOTO(out, mode);
1532
1533                 if (ldlm_is_cbpending(lock) &&
1534                     lock->l_readers == 0 && lock->l_writers == 0)
1535                         GOTO(out, mode);
1536
1537                 if (bits)
1538                         *bits = lock->l_policy_data.l_inodebits.bits;
1539                 mode = lock->l_granted_mode;
1540                 ldlm_lock_addref_internal_nolock(lock, mode);
1541         }
1542
1543         EXIT;
1544
1545 out:
1546         if (lock != NULL) {
1547                 unlock_res_and_lock(lock);
1548                 LDLM_LOCK_PUT(lock);
1549         }
1550         return mode;
1551 }
1552 EXPORT_SYMBOL(ldlm_revalidate_lock_handle);
1553
1554 /** The caller must guarantee that the buffer is large enough. */
1555 int ldlm_fill_lvb(struct ldlm_lock *lock, struct req_capsule *pill,
1556                   enum req_location loc, void *data, int size)
1557 {
1558         void *lvb;
1559         ENTRY;
1560
1561         LASSERT(data != NULL);
1562         LASSERT(size >= 0);
1563
1564         switch (lock->l_lvb_type) {
1565         case LVB_T_OST:
1566                 if (size == sizeof(struct ost_lvb)) {
1567                         if (loc == RCL_CLIENT)
1568                                 lvb = req_capsule_client_swab_get(pill,
1569                                                 &RMF_DLM_LVB,
1570                                                 lustre_swab_ost_lvb);
1571                         else
1572                                 lvb = req_capsule_server_swab_get(pill,
1573                                                 &RMF_DLM_LVB,
1574                                                 lustre_swab_ost_lvb);
1575                         if (unlikely(lvb == NULL)) {
1576                                 LDLM_ERROR(lock, "no LVB");
1577                                 RETURN(-EPROTO);
1578                         }
1579
1580                         memcpy(data, lvb, size);
1581                 } else if (size == sizeof(struct ost_lvb_v1)) {
1582                         struct ost_lvb *olvb = data;
1583
1584                         if (loc == RCL_CLIENT)
1585                                 lvb = req_capsule_client_swab_get(pill,
1586                                                 &RMF_DLM_LVB,
1587                                                 lustre_swab_ost_lvb_v1);
1588                         else
1589                                 lvb = req_capsule_server_sized_swab_get(pill,
1590                                                 &RMF_DLM_LVB, size,
1591                                                 lustre_swab_ost_lvb_v1);
1592                         if (unlikely(lvb == NULL)) {
1593                                 LDLM_ERROR(lock, "no LVB");
1594                                 RETURN(-EPROTO);
1595                         }
1596
1597                         memcpy(data, lvb, size);
1598                         olvb->lvb_mtime_ns = 0;
1599                         olvb->lvb_atime_ns = 0;
1600                         olvb->lvb_ctime_ns = 0;
1601                 } else {
1602                         LDLM_ERROR(lock, "Replied unexpected ost LVB size %d",
1603                                    size);
1604                         RETURN(-EINVAL);
1605                 }
1606                 break;
1607         case LVB_T_LQUOTA:
1608                 if (size == sizeof(struct lquota_lvb)) {
1609                         if (loc == RCL_CLIENT)
1610                                 lvb = req_capsule_client_swab_get(pill,
1611                                                 &RMF_DLM_LVB,
1612                                                 lustre_swab_lquota_lvb);
1613                         else
1614                                 lvb = req_capsule_server_swab_get(pill,
1615                                                 &RMF_DLM_LVB,
1616                                                 lustre_swab_lquota_lvb);
1617                         if (unlikely(lvb == NULL)) {
1618                                 LDLM_ERROR(lock, "no LVB");
1619                                 RETURN(-EPROTO);
1620                         }
1621
1622                         memcpy(data, lvb, size);
1623                 } else {
1624                         LDLM_ERROR(lock, "Replied unexpected lquota LVB size %d",
1625                                    size);
1626                         RETURN(-EINVAL);
1627                 }
1628                 break;
1629         case LVB_T_LAYOUT:
1630                 if (size == 0)
1631                         break;
1632
1633                 if (loc == RCL_CLIENT)
1634                         lvb = req_capsule_client_get(pill, &RMF_DLM_LVB);
1635                 else
1636                         lvb = req_capsule_server_get(pill, &RMF_DLM_LVB);
1637                 if (unlikely(lvb == NULL)) {
1638                         LDLM_ERROR(lock, "no LVB");
1639                         RETURN(-EPROTO);
1640                 }
1641
1642                 memcpy(data, lvb, size);
1643                 break;
1644         default:
1645                 LDLM_ERROR(lock, "Unknown LVB type: %d", lock->l_lvb_type);
1646                 libcfs_debug_dumpstack(NULL);
1647                 RETURN(-EINVAL);
1648         }
1649
1650         RETURN(0);
1651 }
1652
1653 /**
1654  * Create and fill in new LDLM lock with specified properties.
1655  * Returns a referenced lock
1656  */
1657 struct ldlm_lock *ldlm_lock_create(struct ldlm_namespace *ns,
1658                                    const struct ldlm_res_id *res_id,
1659                                    enum ldlm_type type,
1660                                    enum ldlm_mode mode,
1661                                    const struct ldlm_callback_suite *cbs,
1662                                    void *data, __u32 lvb_len,
1663                                    enum lvb_type lvb_type)
1664 {
1665         struct ldlm_lock        *lock;
1666         struct ldlm_resource    *res;
1667         int                     rc;
1668         ENTRY;
1669
1670         res = ldlm_resource_get(ns, NULL, res_id, type, 1);
1671         if (IS_ERR(res))
1672                 RETURN(ERR_CAST(res));
1673
1674         lock = ldlm_lock_new(res);
1675         if (!lock) {
1676                 ldlm_resource_putref(res);
1677                 RETURN(ERR_PTR(-ENOMEM));
1678         }
1679
1680         lock->l_req_mode = mode;
1681         lock->l_ast_data = data;
1682         lock->l_pid = current->pid;
1683         if (ns_is_server(ns))
1684                 ldlm_set_ns_srv(lock);
1685         if (cbs) {
1686                 lock->l_blocking_ast = cbs->lcs_blocking;
1687                 lock->l_completion_ast = cbs->lcs_completion;
1688                 lock->l_glimpse_ast = cbs->lcs_glimpse;
1689         }
1690
1691         switch (type) {
1692         case LDLM_EXTENT:
1693                 rc = ldlm_extent_alloc_lock(lock);
1694                 break;
1695         case LDLM_IBITS:
1696                 rc = ldlm_inodebits_alloc_lock(lock);
1697                 break;
1698         default:
1699                 rc = 0;
1700         }
1701         if (rc)
1702                 GOTO(out, rc);
1703
1704         if (lvb_len) {
1705                 lock->l_lvb_len = lvb_len;
1706                 OBD_ALLOC_LARGE(lock->l_lvb_data, lvb_len);
1707                 if (lock->l_lvb_data == NULL)
1708                         GOTO(out, rc = -ENOMEM);
1709         }
1710
1711         lock->l_lvb_type = lvb_type;
1712         if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_NEW_LOCK))
1713                 GOTO(out, rc = -ENOENT);
1714
1715         RETURN(lock);
1716
1717 out:
1718         ldlm_lock_destroy(lock);
1719         LDLM_LOCK_RELEASE(lock);
1720         RETURN(ERR_PTR(rc));
1721 }
1722
1723 #ifdef HAVE_SERVER_SUPPORT
1724 static enum ldlm_error ldlm_lock_enqueue_helper(struct ldlm_lock *lock,
1725                                              __u64 *flags)
1726 {
1727         struct ldlm_resource *res = lock->l_resource;
1728         enum ldlm_error rc = ELDLM_OK;
1729         LIST_HEAD(rpc_list);
1730         ldlm_processing_policy policy;
1731
1732         ENTRY;
1733
1734         policy = ldlm_get_processing_policy(res);
1735         policy(lock, flags, LDLM_PROCESS_ENQUEUE, &rc, &rpc_list);
1736         if (rc == ELDLM_OK && lock->l_granted_mode != lock->l_req_mode &&
1737             res->lr_type != LDLM_FLOCK)
1738                 rc = ldlm_handle_conflict_lock(lock, flags, &rpc_list);
1739
1740         if (!list_empty(&rpc_list))
1741                 ldlm_discard_bl_list(&rpc_list);
1742
1743         RETURN(rc);
1744 }
1745 #endif
1746
1747 /**
1748  * Enqueue (request) a lock.
1749  *
1750  * Does not block. As a result of enqueue the lock would be put
1751  * into granted or waiting list.
1752  *
1753  * If namespace has intent policy sent and the lock has LDLM_FL_HAS_INTENT flag
1754  * set, skip all the enqueueing and delegate lock processing to intent policy
1755  * function.
1756  */
1757 enum ldlm_error ldlm_lock_enqueue(const struct lu_env *env,
1758                                   struct ldlm_namespace *ns,
1759                                   struct ldlm_lock **lockp,
1760                                   void *cookie, __u64 *flags)
1761 {
1762         struct ldlm_lock *lock = *lockp;
1763         struct ldlm_resource *res;
1764         int local = ns_is_client(ns);
1765         enum ldlm_error rc = ELDLM_OK;
1766         struct ldlm_interval *node = NULL;
1767 #ifdef HAVE_SERVER_SUPPORT
1768         bool reconstruct = false;
1769 #endif
1770         ENTRY;
1771
1772         /* policies are not executed on the client or during replay */
1773         if ((*flags & (LDLM_FL_HAS_INTENT|LDLM_FL_REPLAY)) == LDLM_FL_HAS_INTENT
1774             && !local && ns->ns_policy) {
1775                 rc = ns->ns_policy(env, ns, lockp, cookie, lock->l_req_mode,
1776                                    *flags, NULL);
1777                 if (rc == ELDLM_LOCK_REPLACED) {
1778                         /* The lock that was returned has already been granted,
1779                          * and placed into lockp.  If it's not the same as the
1780                          * one we passed in, then destroy the old one and our
1781                          * work here is done. */
1782                         if (lock != *lockp) {
1783                                 ldlm_lock_destroy(lock);
1784                                 LDLM_LOCK_RELEASE(lock);
1785                         }
1786                         *flags |= LDLM_FL_LOCK_CHANGED;
1787                         RETURN(0);
1788                 } else if (rc != ELDLM_OK &&
1789                            ldlm_is_granted(lock)) {
1790                         LASSERT(*flags & LDLM_FL_RESENT);
1791                         /* It may happen that ns_policy returns an error in
1792                          * resend case, object may be unlinked or just some
1793                          * error occurs. It is unclear if lock reached the
1794                          * client in the original reply, just leave the lock on
1795                          * server, not returning it again to client. Due to
1796                          * LU-6529, the server will not OOM. */
1797                         RETURN(rc);
1798                 } else if (rc != ELDLM_OK ||
1799                            (rc == ELDLM_OK && (*flags & LDLM_FL_INTENT_ONLY))) {
1800                         ldlm_lock_destroy(lock);
1801                         RETURN(rc);
1802                 }
1803         }
1804
1805         if (*flags & LDLM_FL_RESENT) {
1806                 /* Reconstruct LDLM_FL_SRV_ENQ_MASK @flags for reply.
1807                  * Set LOCK_CHANGED always.
1808                  * Check if the lock is granted for BLOCK_GRANTED.
1809                  * Take NO_TIMEOUT from the lock as it is inherited through
1810                  * LDLM_FL_INHERIT_MASK */
1811                 *flags |= LDLM_FL_LOCK_CHANGED;
1812                 if (!ldlm_is_granted(lock))
1813                         *flags |= LDLM_FL_BLOCK_GRANTED;
1814                 *flags |= lock->l_flags & LDLM_FL_NO_TIMEOUT;
1815                 RETURN(ELDLM_OK);
1816         }
1817
1818 #ifdef HAVE_SERVER_SUPPORT
1819         /* For a replaying lock, it might be already in granted list. So
1820          * unlinking the lock will cause the interval node to be freed, we
1821          * have to allocate the interval node early otherwise we can't regrant
1822          * this lock in the future. - jay
1823          *
1824          * The only time the ldlm_resource changes for the ldlm_lock is when
1825          * ldlm_lock_change_resource() is called and that only happens for
1826          * the Lustre client case.
1827          */
1828         if (!local && (*flags & LDLM_FL_REPLAY) &&
1829             lock->l_resource->lr_type == LDLM_EXTENT)
1830                 OBD_SLAB_ALLOC_PTR_GFP(node, ldlm_interval_slab, GFP_NOFS);
1831
1832         reconstruct = !local && lock->l_resource->lr_type == LDLM_FLOCK &&
1833                       !(*flags & LDLM_FL_TEST_LOCK);
1834         if (reconstruct) {
1835                 rc = req_can_reconstruct(cookie, NULL);
1836                 if (rc != 0) {
1837                         if (rc == 1)
1838                                 rc = 0;
1839                         RETURN(rc);
1840                 }
1841         }
1842 #endif
1843         res = lock_res_and_lock(lock);
1844         if (local && ldlm_is_granted(lock)) {
1845                 /* The server returned a blocked lock, but it was granted
1846                  * before we got a chance to actually enqueue it.  We don't
1847                  * need to do anything else. */
1848                 *flags &= ~LDLM_FL_BLOCKED_MASK;
1849                 GOTO(out, rc = ELDLM_OK);
1850         }
1851
1852         ldlm_resource_unlink_lock(lock);
1853         if (res->lr_type == LDLM_EXTENT && lock->l_tree_node == NULL) {
1854                 if (node == NULL) {
1855                         ldlm_lock_destroy_nolock(lock);
1856                         GOTO(out, rc = -ENOMEM);
1857                 }
1858
1859                 INIT_LIST_HEAD(&node->li_group);
1860                 ldlm_interval_attach(node, lock);
1861                 node = NULL;
1862         }
1863
1864         /* Some flags from the enqueue want to make it into the AST, via the
1865          * lock's l_flags. */
1866         if (*flags & LDLM_FL_AST_DISCARD_DATA)
1867                 ldlm_set_ast_discard_data(lock);
1868         if (*flags & LDLM_FL_TEST_LOCK)
1869                 ldlm_set_test_lock(lock);
1870         if (*flags & LDLM_FL_COS_INCOMPAT)
1871                 ldlm_set_cos_incompat(lock);
1872         if (*flags & LDLM_FL_COS_ENABLED)
1873                 ldlm_set_cos_enabled(lock);
1874
1875         /* This distinction between local lock trees is very important; a client
1876          * namespace only has information about locks taken by that client, and
1877          * thus doesn't have enough information to decide for itself if it can
1878          * be granted (below).  In this case, we do exactly what the server
1879          * tells us to do, as dictated by the 'flags'.
1880          *
1881          * We do exactly the same thing during recovery, when the server is
1882          * more or less trusting the clients not to lie.
1883          *
1884          * FIXME (bug 268): Detect obvious lies by checking compatibility in
1885          * granted queue. */
1886         if (local) {
1887                 if (*flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED))
1888                         ldlm_resource_add_lock(res, &res->lr_waiting, lock);
1889                 else
1890                         ldlm_grant_lock(lock, NULL);
1891                 GOTO(out, rc = ELDLM_OK);
1892 #ifdef HAVE_SERVER_SUPPORT
1893         } else if (*flags & LDLM_FL_REPLAY) {
1894                 if (*flags & LDLM_FL_BLOCK_WAIT) {
1895                         ldlm_resource_add_lock(res, &res->lr_waiting, lock);
1896                         GOTO(out, rc = ELDLM_OK);
1897                 } else if (*flags & LDLM_FL_BLOCK_GRANTED) {
1898                         ldlm_grant_lock(lock, NULL);
1899                         GOTO(out, rc = ELDLM_OK);
1900                 }
1901                 /* If no flags, fall through to normal enqueue path. */
1902         }
1903
1904         rc = ldlm_lock_enqueue_helper(lock, flags);
1905         GOTO(out, rc);
1906 #else
1907         } else {
1908                 CERROR("This is client-side-only module, cannot handle "
1909                        "LDLM_NAMESPACE_SERVER resource type lock.\n");
1910                 LBUG();
1911         }
1912 #endif
1913
1914 out:
1915         unlock_res_and_lock(lock);
1916
1917 #ifdef HAVE_SERVER_SUPPORT
1918         if (reconstruct) {
1919                 struct ptlrpc_request *req = cookie;
1920
1921                 tgt_mk_reply_data(NULL, NULL,
1922                                   &req->rq_export->exp_target_data,
1923                                   req, 0, NULL, false, 0);
1924         }
1925 #endif
1926         if (node)
1927                 OBD_SLAB_FREE(node, ldlm_interval_slab, sizeof(*node));
1928         return rc;
1929 }
1930
1931 #ifdef HAVE_SERVER_SUPPORT
1932 /**
1933  * Iterate through all waiting locks on a given resource queue and attempt to
1934  * grant them.
1935  *
1936  * Must be called with resource lock held.
1937  */
1938 int ldlm_reprocess_queue(struct ldlm_resource *res, struct list_head *queue,
1939                          struct list_head *work_list,
1940                          enum ldlm_process_intention intention, __u64 hint)
1941 {
1942         struct list_head *tmp, *pos;
1943         ldlm_processing_policy policy;
1944         __u64 flags;
1945         int rc = LDLM_ITER_CONTINUE;
1946         enum ldlm_error err;
1947         LIST_HEAD(bl_ast_list);
1948
1949         ENTRY;
1950
1951         check_res_locked(res);
1952
1953         policy = ldlm_get_processing_policy(res);
1954         LASSERT(policy);
1955         LASSERT(intention == LDLM_PROCESS_RESCAN ||
1956                 intention == LDLM_PROCESS_RECOVERY);
1957
1958 restart:
1959         list_for_each_safe(tmp, pos, queue) {
1960                 struct ldlm_lock *pending;
1961                 LIST_HEAD(rpc_list);
1962
1963                 pending = list_entry(tmp, struct ldlm_lock, l_res_link);
1964
1965                 CDEBUG(D_INFO, "Reprocessing lock %p\n", pending);
1966
1967                 flags = 0;
1968                 rc = policy(pending, &flags, intention, &err, &rpc_list);
1969                 if (pending->l_granted_mode == pending->l_req_mode ||
1970                     res->lr_type == LDLM_FLOCK) {
1971                         list_splice(&rpc_list, work_list);
1972                 } else {
1973                         list_splice(&rpc_list, &bl_ast_list);
1974                 }
1975                 /*
1976                  * When this is called from recovery done, we always want
1977                  * to scan the whole list no matter what 'rc' is returned.
1978                  */
1979                 if (rc != LDLM_ITER_CONTINUE &&
1980                     intention == LDLM_PROCESS_RESCAN)
1981                         break;
1982         }
1983
1984         if (!list_empty(&bl_ast_list)) {
1985                 unlock_res(res);
1986
1987                 rc = ldlm_run_ast_work(ldlm_res_to_ns(res), &bl_ast_list,
1988                                        LDLM_WORK_BL_AST);
1989
1990                 lock_res(res);
1991                 if (rc == -ERESTART)
1992                         GOTO(restart, rc);
1993         }
1994
1995         if (!list_empty(&bl_ast_list))
1996                 ldlm_discard_bl_list(&bl_ast_list);
1997
1998         RETURN(intention == LDLM_PROCESS_RESCAN ? rc : LDLM_ITER_CONTINUE);
1999 }
2000
2001 /**
2002  * Conflicting locks are detected for a lock to be enqueued, add the lock
2003  * into waiting list and send blocking ASTs to the conflicting locks.
2004  *
2005  * \param[in] lock              The lock to be enqueued.
2006  * \param[out] flags            Lock flags for the lock to be enqueued.
2007  * \param[in] rpc_list          Conflicting locks list.
2008  *
2009  * \retval -ERESTART:   Some lock was instantly canceled while sending
2010  *                      blocking ASTs, caller needs to re-check conflicting
2011  *                      locks.
2012  * \retval -EAGAIN:     Lock was destroyed, caller should return error.
2013  * \reval 0:            Lock is successfully added in waiting list.
2014  */
2015 int ldlm_handle_conflict_lock(struct ldlm_lock *lock, __u64 *flags,
2016                               struct list_head *rpc_list)
2017 {
2018         struct ldlm_resource *res = lock->l_resource;
2019         int rc;
2020         ENTRY;
2021
2022         check_res_locked(res);
2023
2024         /* If either of the compat_queue()s returned failure, then we
2025          * have ASTs to send and must go onto the waiting list.
2026          *
2027          * bug 2322: we used to unlink and re-add here, which was a
2028          * terrible folly -- if we goto restart, we could get
2029          * re-ordered!  Causes deadlock, because ASTs aren't sent! */
2030         if (list_empty(&lock->l_res_link))
2031                 ldlm_resource_add_lock(res, &res->lr_waiting, lock);
2032         unlock_res(res);
2033
2034         rc = ldlm_run_ast_work(ldlm_res_to_ns(res), rpc_list,
2035                                LDLM_WORK_BL_AST);
2036
2037         if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_OST_FAIL_RACE) &&
2038             !ns_is_client(ldlm_res_to_ns(res)))
2039                 class_fail_export(lock->l_export);
2040
2041         if (rc == -ERESTART)
2042                 ldlm_reprocess_all(res, 0);
2043
2044         lock_res(res);
2045         if (rc == -ERESTART) {
2046                 /* 15715: The lock was granted and destroyed after
2047                  * resource lock was dropped. Interval node was freed
2048                  * in ldlm_lock_destroy. Anyway, this always happens
2049                  * when a client is being evicted. So it would be
2050                  * ok to return an error. -jay */
2051                 if (ldlm_is_destroyed(lock))
2052                         RETURN(-EAGAIN);
2053
2054                 /* lock was granted while resource was unlocked. */
2055                 if (ldlm_is_granted(lock)) {
2056                         /* bug 11300: if the lock has been granted,
2057                          * break earlier because otherwise, we will go
2058                          * to restart and ldlm_resource_unlink will be
2059                          * called and it causes the interval node to be
2060                          * freed. Then we will fail at
2061                          * ldlm_extent_add_lock() */
2062                         *flags &= ~LDLM_FL_BLOCKED_MASK;
2063                 }
2064
2065         }
2066         *flags |= LDLM_FL_BLOCK_GRANTED;
2067
2068         RETURN(0);
2069 }
2070
2071 /**
2072  * Discard all AST work items from list.
2073  *
2074  * If for whatever reason we do not want to send ASTs to conflicting locks
2075  * anymore, disassemble the list with this function.
2076  */
2077 void ldlm_discard_bl_list(struct list_head *bl_list)
2078 {
2079         struct ldlm_lock *lock, *tmp;
2080
2081         ENTRY;
2082
2083         list_for_each_entry_safe(lock, tmp, bl_list, l_bl_ast) {
2084                 LASSERT(!list_empty(&lock->l_bl_ast));
2085                 list_del_init(&lock->l_bl_ast);
2086                 ldlm_clear_ast_sent(lock);
2087                 LASSERT(lock->l_bl_ast_run == 0);
2088                 ldlm_clear_blocking_lock(lock);
2089                 LDLM_LOCK_RELEASE(lock);
2090         }
2091         EXIT;
2092 }
2093
2094 /**
2095  * Process a call to blocking AST callback for a lock in ast_work list
2096  */
2097 static int
2098 ldlm_work_bl_ast_lock(struct ptlrpc_request_set *rqset, void *opaq)
2099 {
2100         struct ldlm_cb_set_arg *arg = opaq;
2101         struct ldlm_lock *lock;
2102         struct ldlm_lock_desc d;
2103         struct ldlm_bl_desc bld;
2104         int rc;
2105
2106         ENTRY;
2107
2108         if (list_empty(arg->list))
2109                 RETURN(-ENOENT);
2110
2111         lock = list_entry(arg->list->next, struct ldlm_lock, l_bl_ast);
2112
2113         /* nobody should touch l_bl_ast but some locks in the list may become
2114          * granted after lock convert or COS downgrade, these locks should be
2115          * just skipped here and removed from the list.
2116          */
2117         lock_res_and_lock(lock);
2118         list_del_init(&lock->l_bl_ast);
2119
2120         /* lock is not blocking lock anymore, but was kept in the list because
2121          * it can managed only here.
2122          */
2123         if (!ldlm_is_ast_sent(lock)) {
2124                 unlock_res_and_lock(lock);
2125                 LDLM_LOCK_RELEASE(lock);
2126                 RETURN(0);
2127         }
2128
2129         LASSERT(lock->l_blocking_lock);
2130         ldlm_lock2desc(lock->l_blocking_lock, &d);
2131         /* copy blocking lock ibits in cancel_bits as well,
2132          * new client may use them for lock convert and it is
2133          * important to use new field to convert locks from
2134          * new servers only
2135          */
2136         d.l_policy_data.l_inodebits.cancel_bits =
2137                 lock->l_blocking_lock->l_policy_data.l_inodebits.bits;
2138
2139         /* Blocking lock is being destroyed here but some information about it
2140          * may be needed inside l_blocking_ast() function below,
2141          * e.g. in mdt_blocking_ast(). So save needed data in bl_desc.
2142          */
2143         bld.bl_same_client = lock->l_client_cookie ==
2144                              lock->l_blocking_lock->l_client_cookie;
2145         bld.bl_cos_incompat = ldlm_is_cos_incompat(lock->l_blocking_lock);
2146         arg->bl_desc = &bld;
2147
2148         LASSERT(ldlm_is_ast_sent(lock));
2149         LASSERT(lock->l_bl_ast_run == 0);
2150         lock->l_bl_ast_run++;
2151         ldlm_clear_blocking_lock(lock);
2152         unlock_res_and_lock(lock);
2153
2154         rc = lock->l_blocking_ast(lock, &d, (void *)arg, LDLM_CB_BLOCKING);
2155
2156         LDLM_LOCK_RELEASE(lock);
2157
2158         RETURN(rc);
2159 }
2160
2161 /**
2162  * Process a call to revocation AST callback for a lock in ast_work list
2163  */
2164 static int
2165 ldlm_work_revoke_ast_lock(struct ptlrpc_request_set *rqset, void *opaq)
2166 {
2167         struct ldlm_cb_set_arg *arg = opaq;
2168         struct ldlm_lock_desc   desc;
2169         int                     rc;
2170         struct ldlm_lock       *lock;
2171         ENTRY;
2172
2173         if (list_empty(arg->list))
2174                 RETURN(-ENOENT);
2175
2176         lock = list_entry(arg->list->next, struct ldlm_lock, l_rk_ast);
2177         list_del_init(&lock->l_rk_ast);
2178
2179         /* the desc just pretend to exclusive */
2180         ldlm_lock2desc(lock, &desc);
2181         desc.l_req_mode = LCK_EX;
2182         desc.l_granted_mode = 0;
2183
2184         rc = lock->l_blocking_ast(lock, &desc, (void*)arg, LDLM_CB_BLOCKING);
2185         LDLM_LOCK_RELEASE(lock);
2186
2187         RETURN(rc);
2188 }
2189
2190 /**
2191  * Process a call to glimpse AST callback for a lock in ast_work list
2192  */
2193 int ldlm_work_gl_ast_lock(struct ptlrpc_request_set *rqset, void *opaq)
2194 {
2195         struct ldlm_cb_set_arg          *arg = opaq;
2196         struct ldlm_glimpse_work        *gl_work;
2197         struct ldlm_lock                *lock;
2198         int                              rc = 0;
2199         ENTRY;
2200
2201         if (list_empty(arg->list))
2202                 RETURN(-ENOENT);
2203
2204         gl_work = list_entry(arg->list->next, struct ldlm_glimpse_work,
2205                                  gl_list);
2206         list_del_init(&gl_work->gl_list);
2207
2208         lock = gl_work->gl_lock;
2209
2210         /* transfer the glimpse descriptor to ldlm_cb_set_arg */
2211         arg->gl_desc = gl_work->gl_desc;
2212         arg->gl_interpret_reply = gl_work->gl_interpret_reply;
2213         arg->gl_interpret_data = gl_work->gl_interpret_data;
2214
2215         /* invoke the actual glimpse callback */
2216         rc = lock->l_glimpse_ast(lock, (void *)arg);
2217         if (rc == 0)
2218                 rc = 1; /* update LVB if this is server lock */
2219         else if (rc == -ELDLM_NO_LOCK_DATA)
2220                 ldlm_lvbo_update(lock->l_resource, lock, NULL, 1);
2221
2222         LDLM_LOCK_RELEASE(lock);
2223         if (gl_work->gl_flags & LDLM_GL_WORK_SLAB_ALLOCATED)
2224                 OBD_SLAB_FREE_PTR(gl_work, ldlm_glimpse_work_kmem);
2225         else
2226                 OBD_FREE_PTR(gl_work);
2227
2228         RETURN(rc);
2229 }
2230 #endif
2231
2232 /**
2233  * Process a call to completion AST callback for a lock in ast_work list
2234  */
2235 static int
2236 ldlm_work_cp_ast_lock(struct ptlrpc_request_set *rqset, void *opaq)
2237 {
2238         struct ldlm_cb_set_arg *arg = opaq;
2239         struct ldlm_lock *lock;
2240         ldlm_completion_callback completion_callback;
2241         int rc = 0;
2242
2243         ENTRY;
2244
2245         if (list_empty(arg->list))
2246                 RETURN(-ENOENT);
2247
2248         lock = list_entry(arg->list->next, struct ldlm_lock, l_cp_ast);
2249
2250         /* It's possible to receive a completion AST before we've set
2251          * the l_completion_ast pointer: either because the AST arrived
2252          * before the reply, or simply because there's a small race
2253          * window between receiving the reply and finishing the local
2254          * enqueue. (bug 842)
2255          *
2256          * This can't happen with the blocking_ast, however, because we
2257          * will never call the local blocking_ast until we drop our
2258          * reader/writer reference, which we won't do until we get the
2259          * reply and finish enqueueing. */
2260
2261         /* nobody should touch l_cp_ast */
2262         lock_res_and_lock(lock);
2263         list_del_init(&lock->l_cp_ast);
2264         LASSERT(ldlm_is_cp_reqd(lock));
2265         /* save l_completion_ast since it can be changed by
2266          * mds_intent_policy(), see bug 14225 */
2267         completion_callback = lock->l_completion_ast;
2268         ldlm_clear_cp_reqd(lock);
2269         unlock_res_and_lock(lock);
2270
2271         if (completion_callback != NULL)
2272                 rc = completion_callback(lock, 0, (void *)arg);
2273         LDLM_LOCK_RELEASE(lock);
2274
2275         RETURN(rc);
2276 }
2277
2278 /**
2279  * Process list of locks in need of ASTs being sent.
2280  *
2281  * Used on server to send multiple ASTs together instead of sending one by
2282  * one.
2283  */
2284 int ldlm_run_ast_work(struct ldlm_namespace *ns, struct list_head *rpc_list,
2285                       ldlm_desc_ast_t ast_type)
2286 {
2287         struct ldlm_cb_set_arg *arg;
2288         set_producer_func work_ast_lock;
2289         int rc;
2290
2291         if (list_empty(rpc_list))
2292                 RETURN(0);
2293
2294         OBD_ALLOC_PTR(arg);
2295         if (arg == NULL)
2296                 RETURN(-ENOMEM);
2297
2298         atomic_set(&arg->restart, 0);
2299         arg->list = rpc_list;
2300
2301         switch (ast_type) {
2302         case LDLM_WORK_CP_AST:
2303                 arg->type = LDLM_CP_CALLBACK;
2304                 work_ast_lock = ldlm_work_cp_ast_lock;
2305                 break;
2306 #ifdef HAVE_SERVER_SUPPORT
2307         case LDLM_WORK_BL_AST:
2308                 arg->type = LDLM_BL_CALLBACK;
2309                 work_ast_lock = ldlm_work_bl_ast_lock;
2310                 break;
2311         case LDLM_WORK_REVOKE_AST:
2312                 arg->type = LDLM_BL_CALLBACK;
2313                 work_ast_lock = ldlm_work_revoke_ast_lock;
2314                 break;
2315         case LDLM_WORK_GL_AST:
2316                 arg->type = LDLM_GL_CALLBACK;
2317                 work_ast_lock = ldlm_work_gl_ast_lock;
2318                 break;
2319 #endif
2320         default:
2321                 LBUG();
2322         }
2323
2324         /* We create a ptlrpc request set with flow control extension.
2325          * This request set will use the work_ast_lock function to produce new
2326          * requests and will send a new request each time one completes in order
2327          * to keep the number of requests in flight to ns_max_parallel_ast */
2328         arg->set = ptlrpc_prep_fcset(ns->ns_max_parallel_ast ? : UINT_MAX,
2329                                      work_ast_lock, arg);
2330         if (arg->set == NULL)
2331                 GOTO(out, rc = -ENOMEM);
2332
2333         ptlrpc_set_wait(NULL, arg->set);
2334         ptlrpc_set_destroy(arg->set);
2335
2336         rc = atomic_read(&arg->restart) ? -ERESTART : 0;
2337         GOTO(out, rc);
2338 out:
2339         OBD_FREE_PTR(arg);
2340         return rc;
2341 }
2342
2343 /**
2344  * Try to grant all waiting locks on a resource.
2345  *
2346  * Calls ldlm_reprocess_queue on waiting queue.
2347  *
2348  * Typically called after some resource locks are cancelled to see
2349  * if anything could be granted as a result of the cancellation.
2350  */
2351 static void __ldlm_reprocess_all(struct ldlm_resource *res,
2352                                  enum ldlm_process_intention intention,
2353                                  __u64 hint)
2354 {
2355         LIST_HEAD(rpc_list);
2356 #ifdef HAVE_SERVER_SUPPORT
2357         ldlm_reprocessing_policy reprocess;
2358         struct obd_device *obd;
2359         int rc;
2360
2361         ENTRY;
2362
2363         /* Local lock trees don't get reprocessed. */
2364         if (ns_is_client(ldlm_res_to_ns(res))) {
2365                 EXIT;
2366                 return;
2367         }
2368
2369         /* Disable reprocess during lock replay stage but allow during
2370          * request replay stage.
2371          */
2372         obd = ldlm_res_to_ns(res)->ns_obd;
2373         if (obd->obd_recovering &&
2374             atomic_read(&obd->obd_req_replay_clients) == 0)
2375                 RETURN_EXIT;
2376 restart:
2377         lock_res(res);
2378         reprocess = ldlm_get_reprocessing_policy(res);
2379         reprocess(res, &res->lr_waiting, &rpc_list, intention, hint);
2380         unlock_res(res);
2381
2382         rc = ldlm_run_ast_work(ldlm_res_to_ns(res), &rpc_list,
2383                                LDLM_WORK_CP_AST);
2384         if (rc == -ERESTART) {
2385                 LASSERT(list_empty(&rpc_list));
2386                 hint = 0;
2387                 goto restart;
2388         }
2389 #else
2390         ENTRY;
2391
2392         if (!ns_is_client(ldlm_res_to_ns(res))) {
2393                 CERROR("This is client-side-only module, cannot handle "
2394                        "LDLM_NAMESPACE_SERVER resource type lock.\n");
2395                 LBUG();
2396         }
2397 #endif
2398         EXIT;
2399 }
2400
2401 void ldlm_reprocess_all(struct ldlm_resource *res, __u64 hint)
2402 {
2403         __ldlm_reprocess_all(res, LDLM_PROCESS_RESCAN, hint);
2404 }
2405 EXPORT_SYMBOL(ldlm_reprocess_all);
2406
2407 static int ldlm_reprocess_res(struct cfs_hash *hs, struct cfs_hash_bd *bd,
2408                               struct hlist_node *hnode, void *arg)
2409 {
2410         struct ldlm_resource *res = cfs_hash_object(hs, hnode);
2411
2412         /* This is only called once after recovery done. LU-8306. */
2413         __ldlm_reprocess_all(res, LDLM_PROCESS_RECOVERY, 0);
2414         return 0;
2415 }
2416
2417 /**
2418  * Iterate through all resources on a namespace attempting to grant waiting
2419  * locks.
2420  */
2421 void ldlm_reprocess_recovery_done(struct ldlm_namespace *ns)
2422 {
2423         ENTRY;
2424
2425         if (ns != NULL) {
2426                 cfs_hash_for_each_nolock(ns->ns_rs_hash,
2427                                          ldlm_reprocess_res, NULL, 0);
2428         }
2429         EXIT;
2430 }
2431
2432 /**
2433  * Helper function to call blocking AST for LDLM lock \a lock in a
2434  * "cancelling" mode.
2435  */
2436 void ldlm_cancel_callback(struct ldlm_lock *lock)
2437 {
2438         check_res_locked(lock->l_resource);
2439         if (!ldlm_is_cancel(lock)) {
2440                 ldlm_set_cancel(lock);
2441                 if (lock->l_blocking_ast) {
2442                         unlock_res_and_lock(lock);
2443                         lock->l_blocking_ast(lock, NULL, lock->l_ast_data,
2444                                              LDLM_CB_CANCELING);
2445                         lock_res_and_lock(lock);
2446                 } else {
2447                         LDLM_DEBUG(lock, "no blocking ast");
2448                 }
2449
2450                 /* only canceller can set bl_done bit */
2451                 ldlm_set_bl_done(lock);
2452                 wake_up(&lock->l_waitq);
2453         } else if (!ldlm_is_bl_done(lock)) {
2454                 /* The lock is guaranteed to have been canceled once
2455                  * returning from this function. */
2456                 unlock_res_and_lock(lock);
2457                 wait_event_idle(lock->l_waitq, is_bl_done(lock));
2458                 lock_res_and_lock(lock);
2459         }
2460 }
2461
2462 /**
2463  * Remove skiplist-enabled LDLM lock \a req from granted list
2464  */
2465 void ldlm_unlink_lock_skiplist(struct ldlm_lock *req)
2466 {
2467         if (req->l_resource->lr_type != LDLM_PLAIN &&
2468             req->l_resource->lr_type != LDLM_IBITS)
2469                 return;
2470
2471         list_del_init(&req->l_sl_policy);
2472         list_del_init(&req->l_sl_mode);
2473 }
2474
2475 /**
2476  * Attempts to cancel LDLM lock \a lock that has no reader/writer references.
2477  */
2478 void ldlm_lock_cancel(struct ldlm_lock *lock)
2479 {
2480         struct ldlm_resource *res;
2481         struct ldlm_namespace *ns;
2482         ENTRY;
2483
2484         lock_res_and_lock(lock);
2485
2486         res = lock->l_resource;
2487         ns  = ldlm_res_to_ns(res);
2488
2489         /* Please do not, no matter how tempting, remove this LBUG without
2490          * talking to me first. -phik */
2491         if (lock->l_readers || lock->l_writers) {
2492                 LDLM_ERROR(lock, "lock still has references");
2493                 unlock_res_and_lock(lock);
2494                 LBUG();
2495         }
2496
2497         if (ldlm_is_waited(lock))
2498                 ldlm_del_waiting_lock(lock);
2499
2500         /* Releases cancel callback. */
2501         ldlm_cancel_callback(lock);
2502
2503         /* Yes, second time, just in case it was added again while we were
2504          * running with no res lock in ldlm_cancel_callback */
2505         if (ldlm_is_waited(lock))
2506                 ldlm_del_waiting_lock(lock);
2507
2508         ldlm_resource_unlink_lock(lock);
2509         ldlm_lock_destroy_nolock(lock);
2510
2511         if (ldlm_is_granted(lock))
2512                 ldlm_pool_del(&ns->ns_pool, lock);
2513
2514         /* Make sure we will not be called again for same lock what is possible
2515          * if not to zero out lock->l_granted_mode */
2516         lock->l_granted_mode = LCK_MINMODE;
2517         unlock_res_and_lock(lock);
2518
2519         EXIT;
2520 }
2521 EXPORT_SYMBOL(ldlm_lock_cancel);
2522
2523 /**
2524  * Set opaque data into the lock that only makes sense to upper layer.
2525  */
2526 int ldlm_lock_set_data(const struct lustre_handle *lockh, void *data)
2527 {
2528         struct ldlm_lock *lock = ldlm_handle2lock(lockh);
2529         int rc = -EINVAL;
2530         ENTRY;
2531
2532         if (lock) {
2533                 if (lock->l_ast_data == NULL)
2534                         lock->l_ast_data = data;
2535                 if (lock->l_ast_data == data)
2536                         rc = 0;
2537                 LDLM_LOCK_PUT(lock);
2538         }
2539         RETURN(rc);
2540 }
2541 EXPORT_SYMBOL(ldlm_lock_set_data);
2542
2543 struct export_cl_data {
2544         const struct lu_env     *ecl_env;
2545         struct obd_export       *ecl_exp;
2546         int                     ecl_loop;
2547 };
2548
2549 static void ldlm_cancel_lock_for_export(struct obd_export *exp,
2550                                         struct ldlm_lock *lock,
2551                                         struct export_cl_data *ecl)
2552 {
2553         struct ldlm_resource *res;
2554
2555         res = ldlm_resource_getref(lock->l_resource);
2556
2557         ldlm_lvbo_update(res, lock, NULL, 1);
2558         ldlm_lock_cancel(lock);
2559         if (!exp->exp_obd->obd_stopping)
2560                 ldlm_reprocess_all(res, lock->l_policy_data.l_inodebits.bits);
2561         ldlm_resource_putref(res);
2562
2563         ecl->ecl_loop++;
2564         if ((ecl->ecl_loop & -ecl->ecl_loop) == ecl->ecl_loop) {
2565                 CDEBUG(D_INFO, "Export %p, %d locks cancelled.\n",
2566                        exp, ecl->ecl_loop);
2567         }
2568 }
2569
2570 /**
2571  * Iterator function for ldlm_export_cancel_locks.
2572  * Cancels passed locks.
2573  */
2574 static int
2575 ldlm_cancel_locks_for_export_cb(struct cfs_hash *hs, struct cfs_hash_bd *bd,
2576                                 struct hlist_node *hnode, void *data)
2577
2578 {
2579         struct export_cl_data   *ecl = (struct export_cl_data *)data;
2580         struct obd_export       *exp  = ecl->ecl_exp;
2581         struct ldlm_lock        *lock = cfs_hash_object(hs, hnode);
2582
2583         LDLM_LOCK_GET(lock);
2584         ldlm_cancel_lock_for_export(exp, lock, ecl);
2585         LDLM_LOCK_RELEASE(lock);
2586
2587         return 0;
2588 }
2589
2590 /**
2591  * Cancel all blocked locks for given export.
2592  *
2593  * Typically called on client disconnection/eviction
2594  */
2595 int ldlm_export_cancel_blocked_locks(struct obd_export *exp)
2596 {
2597         struct lu_env env;
2598         struct export_cl_data   ecl = {
2599                 .ecl_exp        = exp,
2600                 .ecl_loop       = 0,
2601         };
2602         int rc;
2603
2604         rc = lu_env_init(&env, LCT_DT_THREAD);
2605         if (rc)
2606                 RETURN(rc);
2607         ecl.ecl_env = &env;
2608
2609         while (!list_empty(&exp->exp_bl_list)) {
2610                 struct ldlm_lock *lock;
2611
2612                 spin_lock_bh(&exp->exp_bl_list_lock);
2613                 if (!list_empty(&exp->exp_bl_list)) {
2614                         lock = list_entry(exp->exp_bl_list.next,
2615                                           struct ldlm_lock, l_exp_list);
2616                         LDLM_LOCK_GET(lock);
2617                         list_del_init(&lock->l_exp_list);
2618                 } else {
2619                         lock = NULL;
2620                 }
2621                 spin_unlock_bh(&exp->exp_bl_list_lock);
2622
2623                 if (lock == NULL)
2624                         break;
2625
2626                 ldlm_cancel_lock_for_export(exp, lock, &ecl);
2627                 LDLM_LOCK_RELEASE(lock);
2628         }
2629
2630         lu_env_fini(&env);
2631
2632         CDEBUG(D_DLMTRACE, "Export %p, canceled %d locks, "
2633                "left on hash table %d.\n", exp, ecl.ecl_loop,
2634                atomic_read(&exp->exp_lock_hash->hs_count));
2635
2636         return ecl.ecl_loop;
2637 }
2638
2639 /**
2640  * Cancel all locks for given export.
2641  *
2642  * Typically called after client disconnection/eviction
2643  */
2644 int ldlm_export_cancel_locks(struct obd_export *exp)
2645 {
2646         struct export_cl_data ecl;
2647         struct lu_env env;
2648         int rc;
2649
2650         rc = lu_env_init(&env, LCT_DT_THREAD);
2651         if (rc)
2652                 RETURN(rc);
2653         ecl.ecl_env = &env;
2654         ecl.ecl_exp = exp;
2655         ecl.ecl_loop = 0;
2656
2657         cfs_hash_for_each_empty(exp->exp_lock_hash,
2658                                 ldlm_cancel_locks_for_export_cb, &ecl);
2659
2660         CDEBUG(D_DLMTRACE, "Export %p, canceled %d locks, "
2661                "left on hash table %d.\n", exp, ecl.ecl_loop,
2662                atomic_read(&exp->exp_lock_hash->hs_count));
2663
2664         if (ecl.ecl_loop > 0 &&
2665             atomic_read(&exp->exp_lock_hash->hs_count) == 0 &&
2666             exp->exp_obd->obd_stopping)
2667                 ldlm_reprocess_recovery_done(exp->exp_obd->obd_namespace);
2668
2669         lu_env_fini(&env);
2670
2671         return ecl.ecl_loop;
2672 }
2673
2674 /**
2675  * Downgrade an PW/EX lock to COS | CR mode.
2676  *
2677  * A lock mode convertion from PW/EX mode to less conflict mode. The
2678  * convertion may fail if lock was canceled before downgrade, but it doesn't
2679  * indicate any problem, because such lock has no reader or writer, and will
2680  * be released soon.
2681  *
2682  * Used by Commit on Sharing (COS) code to force object changes commit in case
2683  * of conflict. Converted lock is considered as new lock and all blocking AST
2684  * things are cleared, so any pending or new blocked lock on that lock will
2685  * cause new call to blocking_ast and force resource object commit.
2686  *
2687  * Also used by layout_change to replace EX lock to CR lock.
2688  *
2689  * \param lock A lock to convert
2690  * \param new_mode new lock mode
2691  */
2692 void ldlm_lock_mode_downgrade(struct ldlm_lock *lock, enum ldlm_mode new_mode)
2693 {
2694 #ifdef HAVE_SERVER_SUPPORT
2695         ENTRY;
2696
2697         LASSERT(new_mode == LCK_COS || new_mode == LCK_CR);
2698
2699         lock_res_and_lock(lock);
2700
2701         if (!(lock->l_granted_mode & (LCK_PW | LCK_EX))) {
2702                 unlock_res_and_lock(lock);
2703
2704                 LASSERT(lock->l_granted_mode == LCK_MINMODE);
2705                 LDLM_DEBUG(lock, "lock was canceled before downgrade");
2706                 RETURN_EXIT;
2707         }
2708
2709         ldlm_resource_unlink_lock(lock);
2710         /*
2711          * Remove the lock from pool as it will be added again in
2712          * ldlm_grant_lock() called below.
2713          */
2714         ldlm_pool_del(&ldlm_lock_to_ns(lock)->ns_pool, lock);
2715
2716         /* Consider downgraded lock as a new lock and clear all states
2717          * related to a previous blocking AST processing.
2718          */
2719         ldlm_clear_blocking_data(lock);
2720
2721         lock->l_req_mode = new_mode;
2722         ldlm_grant_lock(lock, NULL);
2723         unlock_res_and_lock(lock);
2724
2725         ldlm_reprocess_all(lock->l_resource,
2726                            lock->l_policy_data.l_inodebits.bits);
2727
2728         EXIT;
2729 #endif
2730 }
2731 EXPORT_SYMBOL(ldlm_lock_mode_downgrade);
2732
2733 /**
2734  * Print lock with lock handle \a lockh description into debug log.
2735  *
2736  * Used when printing all locks on a resource for debug purposes.
2737  */
2738 void ldlm_lock_dump_handle(int level, const struct lustre_handle *lockh)
2739 {
2740         struct ldlm_lock *lock;
2741
2742         if (!((libcfs_debug | D_ERROR) & level))
2743                 return;
2744
2745         lock = ldlm_handle2lock(lockh);
2746         if (lock == NULL)
2747                 return;
2748
2749         LDLM_DEBUG_LIMIT(level, lock, "###");
2750
2751         LDLM_LOCK_PUT(lock);
2752 }
2753 EXPORT_SYMBOL(ldlm_lock_dump_handle);
2754
2755 /**
2756  * Print lock information with custom message into debug log.
2757  * Helper function.
2758  */
2759 void _ldlm_lock_debug(struct ldlm_lock *lock,
2760                       struct libcfs_debug_msg_data *msgdata,
2761                       const char *fmt, ...)
2762 {
2763         va_list args;
2764         struct obd_export *exp = lock->l_export;
2765         struct ldlm_resource *resource = NULL;
2766         struct va_format vaf;
2767         char *nid = "local";
2768
2769         rcu_read_lock();
2770         resource = rcu_dereference(lock->l_resource);
2771         if (resource && !atomic_inc_not_zero(&resource->lr_refcount))
2772                 resource = NULL;
2773         rcu_read_unlock();
2774
2775         va_start(args, fmt);
2776         vaf.fmt = fmt;
2777         vaf.va = &args;
2778
2779         if (exp && exp->exp_connection) {
2780                 nid = obd_export_nid2str(exp);
2781         } else if (exp && exp->exp_obd != NULL) {
2782                 struct obd_import *imp = exp->exp_obd->u.cli.cl_import;
2783                 nid = obd_import_nid2str(imp);
2784         }
2785
2786         if (resource == NULL) {
2787                 libcfs_debug_msg(msgdata,
2788                                  "%pV ns: \?\? lock: %p/%#llx lrc: %d/%d,%d mode: %s/%s res: \?\? rrc=\?\? type: \?\?\? flags: %#llx nid: %s remote: %#llx expref: %d pid: %u timeout: %lld lvb_type: %d\n",
2789                                  &vaf,
2790                                  lock,
2791                                  lock->l_handle.h_cookie,
2792                                  refcount_read(&lock->l_handle.h_ref),
2793                                  lock->l_readers, lock->l_writers,
2794                                  ldlm_lockname[lock->l_granted_mode],
2795                                  ldlm_lockname[lock->l_req_mode],
2796                                  lock->l_flags, nid,
2797                                  lock->l_remote_handle.cookie,
2798                                  exp ? refcount_read(&exp->exp_handle.h_ref) : -99,
2799                                  lock->l_pid, lock->l_callback_timestamp,
2800                                  lock->l_lvb_type);
2801                 va_end(args);
2802                 return;
2803         }
2804
2805         switch (resource->lr_type) {
2806         case LDLM_EXTENT:
2807                 libcfs_debug_msg(msgdata,
2808                                  "%pV ns: %s lock: %p/%#llx lrc: %d/%d,%d mode: %s/%s res: " DLDLMRES " rrc: %d type: %s [%llu->%llu] (req %llu->%llu) gid %llu flags: %#llx nid: %s remote: %#llx expref: %d pid: %u timeout: %lld lvb_type: %d\n",
2809                                  &vaf,
2810                                  ldlm_lock_to_ns_name(lock), lock,
2811                                  lock->l_handle.h_cookie,
2812                                  refcount_read(&lock->l_handle.h_ref),
2813                                  lock->l_readers, lock->l_writers,
2814                                  ldlm_lockname[lock->l_granted_mode],
2815                                  ldlm_lockname[lock->l_req_mode],
2816                                  PLDLMRES(resource),
2817                                  atomic_read(&resource->lr_refcount),
2818                                  ldlm_typename[resource->lr_type],
2819                                  lock->l_policy_data.l_extent.start,
2820                                  lock->l_policy_data.l_extent.end,
2821                                  lock->l_req_extent.start, lock->l_req_extent.end,
2822                                  lock->l_req_extent.gid,
2823                                  lock->l_flags, nid,
2824                                  lock->l_remote_handle.cookie,
2825                                  exp ? refcount_read(&exp->exp_handle.h_ref) : -99,
2826                                  lock->l_pid, lock->l_callback_timestamp,
2827                                  lock->l_lvb_type);
2828                 break;
2829
2830         case LDLM_FLOCK:
2831                 libcfs_debug_msg(msgdata,
2832                                  "%pV ns: %s lock: %p/%#llx lrc: %d/%d,%d mode: %s/%s res: " DLDLMRES " rrc: %d type: %s pid: %d [%llu->%llu] flags: %#llx nid: %s remote: %#llx expref: %d pid: %u timeout: %lld\n",
2833                                  &vaf,
2834                                  ldlm_lock_to_ns_name(lock), lock,
2835                                  lock->l_handle.h_cookie,
2836                                  refcount_read(&lock->l_handle.h_ref),
2837                                  lock->l_readers, lock->l_writers,
2838                                  ldlm_lockname[lock->l_granted_mode],
2839                                  ldlm_lockname[lock->l_req_mode],
2840                                  PLDLMRES(resource),
2841                                  atomic_read(&resource->lr_refcount),
2842                                  ldlm_typename[resource->lr_type],
2843                                  lock->l_policy_data.l_flock.pid,
2844                                  lock->l_policy_data.l_flock.start,
2845                                  lock->l_policy_data.l_flock.end,
2846                                  lock->l_flags, nid,
2847                                  lock->l_remote_handle.cookie,
2848                                  exp ? refcount_read(&exp->exp_handle.h_ref) : -99,
2849                                  lock->l_pid, lock->l_callback_timestamp);
2850                 break;
2851
2852         case LDLM_IBITS:
2853                 libcfs_debug_msg(msgdata,
2854                                  "%pV ns: %s lock: %p/%#llx lrc: %d/%d,%d mode: %s/%s res: " DLDLMRES " bits %#llx/%#llx rrc: %d type: %s gid %llu flags: %#llx nid: %s remote: %#llx expref: %d pid: %u timeout: %lld lvb_type: %d\n",
2855                                  &vaf,
2856                                  ldlm_lock_to_ns_name(lock),
2857                                  lock, lock->l_handle.h_cookie,
2858                                  refcount_read(&lock->l_handle.h_ref),
2859                                  lock->l_readers, lock->l_writers,
2860                                  ldlm_lockname[lock->l_granted_mode],
2861                                  ldlm_lockname[lock->l_req_mode],
2862                                  PLDLMRES(resource),
2863                                  lock->l_policy_data.l_inodebits.bits,
2864                                  lock->l_policy_data.l_inodebits.try_bits,
2865                                  atomic_read(&resource->lr_refcount),
2866                                  ldlm_typename[resource->lr_type],
2867                                  lock->l_policy_data.l_inodebits.li_gid,
2868                                  lock->l_flags, nid,
2869                                  lock->l_remote_handle.cookie,
2870                                  exp ? refcount_read(&exp->exp_handle.h_ref) : -99,
2871                                  lock->l_pid, lock->l_callback_timestamp,
2872                                  lock->l_lvb_type);
2873                 break;
2874
2875         default:
2876                 libcfs_debug_msg(msgdata,
2877                                  "%pV ns: %s lock: %p/%#llx lrc: %d/%d,%d mode: %s/%s res: " DLDLMRES " rrc: %d type: %s flags: %#llx nid: %s remote: %#llx expref: %d pid: %u timeout: %lld lvb_type: %d\n",
2878                                  &vaf,
2879                                  ldlm_lock_to_ns_name(lock),
2880                                  lock, lock->l_handle.h_cookie,
2881                                  refcount_read(&lock->l_handle.h_ref),
2882                                  lock->l_readers, lock->l_writers,
2883                                  ldlm_lockname[lock->l_granted_mode],
2884                                  ldlm_lockname[lock->l_req_mode],
2885                                  PLDLMRES(resource),
2886                                  atomic_read(&resource->lr_refcount),
2887                                  ldlm_typename[resource->lr_type],
2888                                  lock->l_flags, nid,
2889                                  lock->l_remote_handle.cookie,
2890                                  exp ? refcount_read(&exp->exp_handle.h_ref) : -99,
2891                                  lock->l_pid, lock->l_callback_timestamp,
2892                                  lock->l_lvb_type);
2893                 break;
2894         }
2895         va_end(args);
2896         ldlm_resource_putref(resource);
2897 }
2898 EXPORT_SYMBOL(_ldlm_lock_debug);