Whamcloud - gitweb
LU-4801 ldlm: discard l_lock from struct ldlm_lock.
[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  * Lustre is a trademark of Sun Microsystems, Inc.
31  *
32  * lustre/ldlm/ldlm_lock.c
33  *
34  * Author: Peter Braam <braam@clusterfs.com>
35  * Author: Phil Schwan <phil@clusterfs.com>
36  */
37
38 #define DEBUG_SUBSYSTEM S_LDLM
39
40 #include <libcfs/libcfs.h>
41
42 #include <lustre_swab.h>
43 #include <obd_class.h>
44
45 #include "ldlm_internal.h"
46
47 struct kmem_cache *ldlm_glimpse_work_kmem;
48 EXPORT_SYMBOL(ldlm_glimpse_work_kmem);
49
50 /* lock types */
51 char *ldlm_lockname[] = {
52         [0] = "--",
53         [LCK_EX] = "EX",
54         [LCK_PW] = "PW",
55         [LCK_PR] = "PR",
56         [LCK_CW] = "CW",
57         [LCK_CR] = "CR",
58         [LCK_NL] = "NL",
59         [LCK_GROUP] = "GROUP",
60         [LCK_COS] = "COS"
61 };
62 EXPORT_SYMBOL(ldlm_lockname);
63
64 char *ldlm_typename[] = {
65         [LDLM_PLAIN] = "PLN",
66         [LDLM_EXTENT] = "EXT",
67         [LDLM_FLOCK] = "FLK",
68         [LDLM_IBITS] = "IBT",
69 };
70
71 static ldlm_policy_wire_to_local_t ldlm_policy_wire_to_local[] = {
72         [LDLM_PLAIN - LDLM_MIN_TYPE]  = ldlm_plain_policy_wire_to_local,
73         [LDLM_EXTENT - LDLM_MIN_TYPE] = ldlm_extent_policy_wire_to_local,
74         [LDLM_FLOCK - LDLM_MIN_TYPE]  = ldlm_flock_policy_wire_to_local,
75         [LDLM_IBITS - LDLM_MIN_TYPE]  = ldlm_ibits_policy_wire_to_local,
76 };
77
78 static ldlm_policy_local_to_wire_t ldlm_policy_local_to_wire[] = {
79         [LDLM_PLAIN - LDLM_MIN_TYPE]  = ldlm_plain_policy_local_to_wire,
80         [LDLM_EXTENT - LDLM_MIN_TYPE] = ldlm_extent_policy_local_to_wire,
81         [LDLM_FLOCK - LDLM_MIN_TYPE]  = ldlm_flock_policy_local_to_wire,
82         [LDLM_IBITS - LDLM_MIN_TYPE]  = ldlm_ibits_policy_local_to_wire,
83 };
84
85 /**
86  * Converts lock policy from local format to on the wire lock_desc format
87  */
88 void ldlm_convert_policy_to_wire(enum ldlm_type type,
89                                  const union ldlm_policy_data *lpolicy,
90                                  union ldlm_wire_policy_data *wpolicy)
91 {
92         ldlm_policy_local_to_wire_t convert;
93
94         convert = ldlm_policy_local_to_wire[type - LDLM_MIN_TYPE];
95
96         convert(lpolicy, wpolicy);
97 }
98
99 /**
100  * Converts lock policy from on the wire lock_desc format to local format
101  */
102 void ldlm_convert_policy_to_local(struct obd_export *exp, enum ldlm_type type,
103                                   const union ldlm_wire_policy_data *wpolicy,
104                                   union ldlm_policy_data *lpolicy)
105 {
106         ldlm_policy_wire_to_local_t convert;
107
108         convert = ldlm_policy_wire_to_local[type - LDLM_MIN_TYPE];
109
110         convert(wpolicy, lpolicy);
111 }
112
113 const char *ldlm_it2str(enum ldlm_intent_flags it)
114 {
115         switch (it) {
116         case IT_OPEN:
117                 return "open";
118         case IT_CREAT:
119                 return "creat";
120         case (IT_OPEN | IT_CREAT):
121                 return "open|creat";
122         case IT_READDIR:
123                 return "readdir";
124         case IT_GETATTR:
125                 return "getattr";
126         case IT_LOOKUP:
127                 return "lookup";
128         case IT_GETXATTR:
129                 return "getxattr";
130         case IT_LAYOUT:
131                 return "layout";
132         default:
133                 CERROR("Unknown intent 0x%08x\n", it);
134                 return "UNKNOWN";
135         }
136 }
137 EXPORT_SYMBOL(ldlm_it2str);
138
139 #ifdef HAVE_SERVER_SUPPORT
140 static ldlm_processing_policy ldlm_processing_policy_table[] = {
141         [LDLM_PLAIN]    = ldlm_process_plain_lock,
142         [LDLM_EXTENT]   = ldlm_process_extent_lock,
143         [LDLM_FLOCK]    = ldlm_process_flock_lock,
144         [LDLM_IBITS]    = ldlm_process_inodebits_lock,
145 };
146
147 ldlm_processing_policy ldlm_get_processing_policy(struct ldlm_resource *res)
148 {
149         return ldlm_processing_policy_table[res->lr_type];
150 }
151 EXPORT_SYMBOL(ldlm_get_processing_policy);
152
153 static ldlm_reprocessing_policy ldlm_reprocessing_policy_table[] = {
154         [LDLM_PLAIN]    = ldlm_reprocess_queue,
155         [LDLM_EXTENT]   = ldlm_reprocess_queue,
156         [LDLM_FLOCK]    = ldlm_reprocess_queue,
157         [LDLM_IBITS]    = ldlm_reprocess_inodebits_queue,
158 };
159
160 ldlm_reprocessing_policy ldlm_get_reprocessing_policy(struct ldlm_resource *res)
161 {
162         return ldlm_reprocessing_policy_table[res->lr_type];
163 }
164
165 #endif /* HAVE_SERVER_SUPPORT */
166
167 void ldlm_register_intent(struct ldlm_namespace *ns, ldlm_res_policy arg)
168 {
169         ns->ns_policy = arg;
170 }
171 EXPORT_SYMBOL(ldlm_register_intent);
172
173 /*
174  * REFCOUNTED LOCK OBJECTS
175  */
176
177
178 /**
179  * Get a reference on a lock.
180  *
181  * Lock refcounts, during creation:
182  *   - one special one for allocation, dec'd only once in destroy
183  *   - one for being a lock that's in-use
184  *   - one for the addref associated with a new lock
185  */
186 struct ldlm_lock *ldlm_lock_get(struct ldlm_lock *lock)
187 {
188         refcount_inc(&lock->l_handle.h_ref);
189         return lock;
190 }
191 EXPORT_SYMBOL(ldlm_lock_get);
192
193 static void lock_handle_free(struct rcu_head *rcu)
194 {
195         struct ldlm_lock *lock = container_of(rcu, struct ldlm_lock,
196                                               l_handle.h_rcu);
197
198         OBD_FREE_PRE(lock, sizeof(*lock), "slab-freed");
199         kmem_cache_free(ldlm_lock_slab, lock);
200 }
201
202 /**
203  * Release lock reference.
204  *
205  * Also frees the lock if it was last reference.
206  */
207 void ldlm_lock_put(struct ldlm_lock *lock)
208 {
209         ENTRY;
210
211         LASSERT(lock->l_resource != LP_POISON);
212         LASSERT(refcount_read(&lock->l_handle.h_ref) > 0);
213         if (refcount_dec_and_test(&lock->l_handle.h_ref)) {
214                 struct ldlm_resource *res;
215
216                 LDLM_DEBUG(lock,
217                            "final lock_put on destroyed lock, freeing it.");
218
219                 res = lock->l_resource;
220                 LASSERT(ldlm_is_destroyed(lock));
221                 LASSERT(list_empty(&lock->l_exp_list));
222                 LASSERT(list_empty(&lock->l_res_link));
223                 LASSERT(list_empty(&lock->l_pending_chain));
224
225                 lprocfs_counter_decr(ldlm_res_to_ns(res)->ns_stats,
226                                      LDLM_NSS_LOCKS);
227                 lu_ref_del(&res->lr_reference, "lock", lock);
228                 if (lock->l_export) {
229                         class_export_lock_put(lock->l_export, lock);
230                         lock->l_export = NULL;
231                 }
232
233                 if (lock->l_lvb_data != NULL)
234                         OBD_FREE_LARGE(lock->l_lvb_data, lock->l_lvb_len);
235
236                 if (res->lr_type == LDLM_EXTENT) {
237                         ldlm_interval_free(ldlm_interval_detach(lock));
238                 } else if (res->lr_type == LDLM_IBITS) {
239                         if (lock->l_ibits_node != NULL)
240                                 OBD_SLAB_FREE_PTR(lock->l_ibits_node,
241                                                   ldlm_inodebits_slab);
242                 }
243                 ldlm_resource_putref(res);
244                 lock->l_resource = NULL;
245                 lu_ref_fini(&lock->l_reference);
246                 call_rcu(&lock->l_handle.h_rcu, lock_handle_free);
247         }
248
249         EXIT;
250 }
251 EXPORT_SYMBOL(ldlm_lock_put);
252
253 /**
254  * Removes LDLM lock \a lock from LRU. Assumes LRU is already locked.
255  */
256 int ldlm_lock_remove_from_lru_nolock(struct ldlm_lock *lock)
257 {
258         int rc = 0;
259         if (!list_empty(&lock->l_lru)) {
260                 struct ldlm_namespace *ns = ldlm_lock_to_ns(lock);
261
262                 LASSERT(lock->l_resource->lr_type != LDLM_FLOCK);
263                 if (ns->ns_last_pos == &lock->l_lru)
264                         ns->ns_last_pos = lock->l_lru.prev;
265                 list_del_init(&lock->l_lru);
266                 LASSERT(ns->ns_nr_unused > 0);
267                 ns->ns_nr_unused--;
268                 rc = 1;
269         }
270         return rc;
271 }
272
273 /**
274  * Removes LDLM lock \a lock from LRU. Obtains the LRU lock first.
275  *
276  * If \a last_use is non-zero, it will remove the lock from LRU only if
277  * it matches lock's l_last_used.
278  *
279  * \retval 0 if \a last_use is set, the lock is not in LRU list or \a last_use
280  *           doesn't match lock's l_last_used;
281  *           otherwise, the lock hasn't been in the LRU list.
282  * \retval 1 the lock was in LRU list and removed.
283  */
284 int ldlm_lock_remove_from_lru_check(struct ldlm_lock *lock, ktime_t last_use)
285 {
286         struct ldlm_namespace *ns = ldlm_lock_to_ns(lock);
287         int rc = 0;
288
289         ENTRY;
290         if (ldlm_is_ns_srv(lock)) {
291                 LASSERT(list_empty(&lock->l_lru));
292                 RETURN(0);
293         }
294
295         spin_lock(&ns->ns_lock);
296         if (!ktime_compare(last_use, ktime_set(0, 0)) ||
297             !ktime_compare(last_use, lock->l_last_used))
298                 rc = ldlm_lock_remove_from_lru_nolock(lock);
299         spin_unlock(&ns->ns_lock);
300
301         RETURN(rc);
302 }
303
304 /**
305  * Adds LDLM lock \a lock to namespace LRU. Assumes LRU is already locked.
306  */
307 void ldlm_lock_add_to_lru_nolock(struct ldlm_lock *lock)
308 {
309         struct ldlm_namespace *ns = ldlm_lock_to_ns(lock);
310
311         lock->l_last_used = ktime_get();
312         LASSERT(list_empty(&lock->l_lru));
313         LASSERT(lock->l_resource->lr_type != LDLM_FLOCK);
314         list_add_tail(&lock->l_lru, &ns->ns_unused_list);
315         LASSERT(ns->ns_nr_unused >= 0);
316         ns->ns_nr_unused++;
317 }
318
319 /**
320  * Adds LDLM lock \a lock to namespace LRU. Obtains necessary LRU locks
321  * first.
322  */
323 void ldlm_lock_add_to_lru(struct ldlm_lock *lock)
324 {
325         struct ldlm_namespace *ns = ldlm_lock_to_ns(lock);
326
327         ENTRY;
328         spin_lock(&ns->ns_lock);
329         ldlm_lock_add_to_lru_nolock(lock);
330         spin_unlock(&ns->ns_lock);
331         EXIT;
332 }
333
334 /**
335  * Moves LDLM lock \a lock that is already in namespace LRU to the tail of
336  * the LRU. Performs necessary LRU locking
337  */
338 void ldlm_lock_touch_in_lru(struct ldlm_lock *lock)
339 {
340         struct ldlm_namespace *ns = ldlm_lock_to_ns(lock);
341
342         ENTRY;
343         if (ldlm_is_ns_srv(lock)) {
344                 LASSERT(list_empty(&lock->l_lru));
345                 EXIT;
346                 return;
347         }
348
349         spin_lock(&ns->ns_lock);
350         if (!list_empty(&lock->l_lru)) {
351                 ldlm_lock_remove_from_lru_nolock(lock);
352                 ldlm_lock_add_to_lru_nolock(lock);
353         }
354         spin_unlock(&ns->ns_lock);
355         EXIT;
356 }
357
358 /**
359  * Helper to destroy a locked lock.
360  *
361  * Used by ldlm_lock_destroy and ldlm_lock_destroy_nolock
362  * Must be called with l_lock and lr_lock held.
363  *
364  * Does not actually free the lock data, but rather marks the lock as
365  * destroyed by setting l_destroyed field in the lock to 1.  Destroys a
366  * handle->lock association too, so that the lock can no longer be found
367  * and removes the lock from LRU list.  Actual lock freeing occurs when
368  * last lock reference goes away.
369  *
370  * Original comment (of some historical value):
371  * This used to have a 'strict' flag, which recovery would use to mark an
372  * in-use lock as needing-to-die.  Lest I am ever tempted to put it back, I
373  * shall explain why it's gone: with the new hash table scheme, once you call
374  * ldlm_lock_destroy, you can never drop your final references on this lock.
375  * Because it's not in the hash table anymore.  -phil
376  */
377 static int ldlm_lock_destroy_internal(struct ldlm_lock *lock)
378 {
379         ENTRY;
380
381         if (lock->l_readers || lock->l_writers) {
382                 LDLM_ERROR(lock, "lock still has references");
383                 LBUG();
384         }
385
386         if (!list_empty(&lock->l_res_link)) {
387                 LDLM_ERROR(lock, "lock still on resource");
388                 LBUG();
389         }
390
391         if (ldlm_is_destroyed(lock)) {
392                 LASSERT(list_empty(&lock->l_lru));
393                 EXIT;
394                 return 0;
395         }
396         ldlm_set_destroyed(lock);
397
398         if (lock->l_export && lock->l_export->exp_lock_hash) {
399                 /* NB: it's safe to call cfs_hash_del() even lock isn't
400                  * in exp_lock_hash. */
401                 /* In the function below, .hs_keycmp resolves to
402                  * ldlm_export_lock_keycmp() */
403                 /* coverity[overrun-buffer-val] */
404                 cfs_hash_del(lock->l_export->exp_lock_hash,
405                              &lock->l_remote_handle, &lock->l_exp_hash);
406         }
407
408         ldlm_lock_remove_from_lru(lock);
409         class_handle_unhash(&lock->l_handle);
410
411         EXIT;
412         return 1;
413 }
414
415 /**
416  * Destroys a LDLM lock \a lock. Performs necessary locking first.
417  */
418 void ldlm_lock_destroy(struct ldlm_lock *lock)
419 {
420         int first;
421         ENTRY;
422         lock_res_and_lock(lock);
423         first = ldlm_lock_destroy_internal(lock);
424         unlock_res_and_lock(lock);
425
426         /* drop reference from hashtable only for first destroy */
427         if (first) {
428                 lu_ref_del(&lock->l_reference, "hash", lock);
429                 LDLM_LOCK_RELEASE(lock);
430         }
431         EXIT;
432 }
433
434 /**
435  * Destroys a LDLM lock \a lock that is already locked.
436  */
437 void ldlm_lock_destroy_nolock(struct ldlm_lock *lock)
438 {
439         int first;
440         ENTRY;
441         first = ldlm_lock_destroy_internal(lock);
442         /* drop reference from hashtable only for first destroy */
443         if (first) {
444                 lu_ref_del(&lock->l_reference, "hash", lock);
445                 LDLM_LOCK_RELEASE(lock);
446         }
447         EXIT;
448 }
449
450 static const char lock_handle_owner[] = "ldlm";
451
452 /**
453  *
454  * Allocate and initialize new lock structure.
455  *
456  * usage: pass in a resource on which you have done ldlm_resource_get
457  *        new lock will take over the refcount.
458  * returns: lock with refcount 2 - one for current caller and one for remote
459  */
460 static struct ldlm_lock *ldlm_lock_new(struct ldlm_resource *resource)
461 {
462         struct ldlm_lock *lock;
463         ENTRY;
464
465         if (resource == NULL)
466                 LBUG();
467
468         OBD_SLAB_ALLOC_PTR_GFP(lock, ldlm_lock_slab, GFP_NOFS);
469         if (lock == NULL)
470                 RETURN(NULL);
471
472         RCU_INIT_POINTER(lock->l_resource, resource);
473         lu_ref_add(&resource->lr_reference, "lock", lock);
474
475         refcount_set(&lock->l_handle.h_ref, 2);
476         INIT_LIST_HEAD(&lock->l_res_link);
477         INIT_LIST_HEAD(&lock->l_lru);
478         INIT_LIST_HEAD(&lock->l_pending_chain);
479         INIT_LIST_HEAD(&lock->l_bl_ast);
480         INIT_LIST_HEAD(&lock->l_cp_ast);
481         INIT_LIST_HEAD(&lock->l_rk_ast);
482         init_waitqueue_head(&lock->l_waitq);
483         lock->l_blocking_lock = NULL;
484         INIT_LIST_HEAD(&lock->l_sl_mode);
485         INIT_LIST_HEAD(&lock->l_sl_policy);
486         INIT_HLIST_NODE(&lock->l_exp_hash);
487         INIT_HLIST_NODE(&lock->l_exp_flock_hash);
488
489         lprocfs_counter_incr(ldlm_res_to_ns(resource)->ns_stats,
490                              LDLM_NSS_LOCKS);
491         INIT_HLIST_NODE(&lock->l_handle.h_link);
492         class_handle_hash(&lock->l_handle, lock_handle_owner);
493
494         lu_ref_init(&lock->l_reference);
495         lu_ref_add(&lock->l_reference, "hash", lock);
496         lock->l_callback_timestamp = 0;
497         lock->l_activity = 0;
498
499 #if LUSTRE_TRACKS_LOCK_EXP_REFS
500         INIT_LIST_HEAD(&lock->l_exp_refs_link);
501         lock->l_exp_refs_nr = 0;
502         lock->l_exp_refs_target = NULL;
503 #endif
504         INIT_LIST_HEAD(&lock->l_exp_list);
505
506         RETURN(lock);
507 }
508
509 /**
510  * Moves LDLM lock \a lock to another resource.
511  * This is used on client when server returns some other lock than requested
512  * (typically as a result of intent operation)
513  */
514 int ldlm_lock_change_resource(struct ldlm_namespace *ns, struct ldlm_lock *lock,
515                               const struct ldlm_res_id *new_resid)
516 {
517         struct ldlm_resource *oldres;
518         struct ldlm_resource *newres;
519         int type;
520         ENTRY;
521
522         LASSERT(ns_is_client(ns));
523
524         oldres = lock_res_and_lock(lock);
525         if (memcmp(new_resid, &oldres->lr_name,
526                    sizeof(oldres->lr_name)) == 0) {
527                 /* Nothing to do */
528                 unlock_res_and_lock(lock);
529                 RETURN(0);
530         }
531
532         LASSERT(new_resid->name[0] != 0);
533
534         /* This function assumes that the lock isn't on any lists */
535         LASSERT(list_empty(&lock->l_res_link));
536
537         type = oldres->lr_type;
538         unlock_res_and_lock(lock);
539
540         newres = ldlm_resource_get(ns, NULL, new_resid, type, 1);
541         if (IS_ERR(newres))
542                 RETURN(PTR_ERR(newres));
543
544         lu_ref_add(&newres->lr_reference, "lock", lock);
545         /*
546          * To flip the lock from the old to the new resource, oldres
547          * and newres have to be locked. Resource spin-locks are taken
548          * in the memory address order to avoid dead-locks.
549          * As this is the only circumstance where ->l_resource
550          * can change, and this cannot race with itself, it is safe
551          * to access lock->l_resource without being careful about locking.
552          */
553         oldres = lock->l_resource;
554         if (oldres < newres) {
555                 lock_res(oldres);
556                 lock_res_nested(newres, LRT_NEW);
557         } else {
558                 lock_res(newres);
559                 lock_res_nested(oldres, LRT_NEW);
560         }
561         LASSERT(memcmp(new_resid, &oldres->lr_name,
562                        sizeof oldres->lr_name) != 0);
563         rcu_assign_pointer(lock->l_resource, newres);
564         unlock_res(oldres);
565         unlock_res(newres);
566
567         /* ...and the flowers are still standing! */
568         lu_ref_del(&oldres->lr_reference, "lock", lock);
569         ldlm_resource_putref(oldres);
570
571         RETURN(0);
572 }
573
574 /** \defgroup ldlm_handles LDLM HANDLES
575  * Ways to get hold of locks without any addresses.
576  * @{
577  */
578
579 /**
580  * Fills in handle for LDLM lock \a lock into supplied \a lockh
581  * Does not take any references.
582  */
583 void ldlm_lock2handle(const struct ldlm_lock *lock, struct lustre_handle *lockh)
584 {
585         lockh->cookie = lock->l_handle.h_cookie;
586 }
587 EXPORT_SYMBOL(ldlm_lock2handle);
588
589 /**
590  * Obtain a lock reference by handle.
591  *
592  * if \a flags: atomically get the lock and set the flags.
593  *              Return NULL if flag already set
594  */
595 struct ldlm_lock *__ldlm_handle2lock(const struct lustre_handle *handle,
596                                      __u64 flags)
597 {
598         struct ldlm_lock *lock;
599         ENTRY;
600
601         LASSERT(handle);
602
603         if (!lustre_handle_is_used(handle))
604                 RETURN(NULL);
605
606         lock = class_handle2object(handle->cookie, lock_handle_owner);
607         if (lock == NULL)
608                 RETURN(NULL);
609
610         if (lock->l_export != NULL && lock->l_export->exp_failed) {
611                 CDEBUG(D_INFO, "lock export failed: lock %p, exp %p\n",
612                        lock, lock->l_export);
613                 LDLM_LOCK_PUT(lock);
614                 RETURN(NULL);
615         }
616
617         /* It's unlikely but possible that someone marked the lock as
618          * destroyed after we did handle2object on it */
619         if ((flags == 0) && !ldlm_is_destroyed(lock)) {
620                 lu_ref_add_atomic(&lock->l_reference, "handle", lock);
621                 RETURN(lock);
622         }
623
624         lock_res_and_lock(lock);
625
626         LASSERT(lock->l_resource != NULL);
627
628         lu_ref_add_atomic(&lock->l_reference, "handle", lock);
629         if (unlikely(ldlm_is_destroyed(lock))) {
630                 unlock_res_and_lock(lock);
631                 CDEBUG(D_INFO, "lock already destroyed: lock %p\n", lock);
632                 LDLM_LOCK_PUT(lock);
633                 RETURN(NULL);
634         }
635
636         /* If we're setting flags, make sure none of them are already set. */
637         if (flags != 0) {
638                 if ((lock->l_flags & flags) != 0) {
639                         unlock_res_and_lock(lock);
640                         LDLM_LOCK_PUT(lock);
641                         RETURN(NULL);
642                 }
643
644                 lock->l_flags |= flags;
645         }
646
647         unlock_res_and_lock(lock);
648         RETURN(lock);
649 }
650 EXPORT_SYMBOL(__ldlm_handle2lock);
651 /** @} ldlm_handles */
652
653 /**
654  * Fill in "on the wire" representation for given LDLM lock into supplied
655  * lock descriptor \a desc structure.
656  */
657 void ldlm_lock2desc(struct ldlm_lock *lock, struct ldlm_lock_desc *desc)
658 {
659         ldlm_res2desc(lock->l_resource, &desc->l_resource);
660         desc->l_req_mode = lock->l_req_mode;
661         desc->l_granted_mode = lock->l_granted_mode;
662         ldlm_convert_policy_to_wire(lock->l_resource->lr_type,
663                                     &lock->l_policy_data,
664                                     &desc->l_policy_data);
665 }
666
667 /**
668  * Add a lock to list of conflicting locks to send AST to.
669  *
670  * Only add if we have not sent a blocking AST to the lock yet.
671  */
672 static void ldlm_add_bl_work_item(struct ldlm_lock *lock, struct ldlm_lock *new,
673                                   struct list_head *work_list)
674 {
675         if (!ldlm_is_ast_sent(lock)) {
676                 LDLM_DEBUG(lock, "lock incompatible; sending blocking AST.");
677                 ldlm_set_ast_sent(lock);
678                 /* If the enqueuing client said so, tell the AST recipient to
679                  * discard dirty data, rather than writing back. */
680                 if (ldlm_is_ast_discard_data(new))
681                         ldlm_set_discard_data(lock);
682
683                 /* Lock can be converted from a blocking state back to granted
684                  * after lock convert or COS downgrade but still be in an
685                  * older bl_list because it is controlled only by
686                  * ldlm_work_bl_ast_lock(), let it be processed there.
687                  */
688                 if (list_empty(&lock->l_bl_ast)) {
689                         list_add(&lock->l_bl_ast, work_list);
690                         LDLM_LOCK_GET(lock);
691                 }
692                 LASSERT(lock->l_blocking_lock == NULL);
693                 lock->l_blocking_lock = LDLM_LOCK_GET(new);
694         }
695 }
696
697 /**
698  * Add a lock to list of just granted locks to send completion AST to.
699  */
700 static void ldlm_add_cp_work_item(struct ldlm_lock *lock,
701                                   struct list_head *work_list)
702 {
703         if (!ldlm_is_cp_reqd(lock)) {
704                 ldlm_set_cp_reqd(lock);
705                 LDLM_DEBUG(lock, "lock granted; sending completion AST.");
706                 LASSERT(list_empty(&lock->l_cp_ast));
707                 list_add(&lock->l_cp_ast, work_list);
708                 LDLM_LOCK_GET(lock);
709         }
710 }
711
712 /**
713  * Aggregator function to add AST work items into a list. Determines
714  * what sort of an AST work needs to be done and calls the proper
715  * adding function.
716  * Must be called with lr_lock held.
717  */
718 void ldlm_add_ast_work_item(struct ldlm_lock *lock, struct ldlm_lock *new,
719                             struct list_head *work_list)
720 {
721         ENTRY;
722         check_res_locked(lock->l_resource);
723         if (new)
724                 ldlm_add_bl_work_item(lock, new, work_list);
725         else
726                 ldlm_add_cp_work_item(lock, work_list);
727         EXIT;
728 }
729
730 /**
731  * Add specified reader/writer reference to LDLM lock with handle \a lockh.
732  * r/w reference type is determined by \a mode
733  * Calls ldlm_lock_addref_internal.
734  */
735 void ldlm_lock_addref(const struct lustre_handle *lockh, enum ldlm_mode mode)
736 {
737         struct ldlm_lock *lock;
738
739         lock = ldlm_handle2lock(lockh);
740         LASSERTF(lock != NULL, "Non-existing lock: %#llx\n", lockh->cookie);
741         ldlm_lock_addref_internal(lock, mode);
742         LDLM_LOCK_PUT(lock);
743 }
744 EXPORT_SYMBOL(ldlm_lock_addref);
745
746 /**
747  * Helper function.
748  * Add specified reader/writer reference to LDLM lock \a lock.
749  * r/w reference type is determined by \a mode
750  * Removes lock from LRU if it is there.
751  * Assumes the LDLM lock is already locked.
752  */
753 void ldlm_lock_addref_internal_nolock(struct ldlm_lock *lock,
754                                       enum ldlm_mode mode)
755 {
756         ldlm_lock_remove_from_lru(lock);
757         if (mode & (LCK_NL | LCK_CR | LCK_PR)) {
758                 lock->l_readers++;
759                 lu_ref_add_atomic(&lock->l_reference, "reader", lock);
760         }
761         if (mode & (LCK_EX | LCK_CW | LCK_PW | LCK_GROUP | LCK_COS)) {
762                 lock->l_writers++;
763                 lu_ref_add_atomic(&lock->l_reference, "writer", lock);
764         }
765         LDLM_LOCK_GET(lock);
766         lu_ref_add_atomic(&lock->l_reference, "user", lock);
767         LDLM_DEBUG(lock, "ldlm_lock_addref(%s)", ldlm_lockname[mode]);
768 }
769
770 /**
771  * Attempts to add reader/writer reference to a lock with handle \a lockh, and
772  * fails if lock is already LDLM_FL_CBPENDING or destroyed.
773  *
774  * \retval 0 success, lock was addref-ed
775  *
776  * \retval -EAGAIN lock is being canceled.
777  */
778 int ldlm_lock_addref_try(const struct lustre_handle *lockh, enum ldlm_mode mode)
779 {
780         struct ldlm_lock *lock;
781         int               result;
782
783         result = -EAGAIN;
784         lock = ldlm_handle2lock(lockh);
785         if (lock != NULL) {
786                 lock_res_and_lock(lock);
787                 if (lock->l_readers != 0 || lock->l_writers != 0 ||
788                     !ldlm_is_cbpending(lock)) {
789                         ldlm_lock_addref_internal_nolock(lock, mode);
790                         result = 0;
791                 }
792                 unlock_res_and_lock(lock);
793                 LDLM_LOCK_PUT(lock);
794         }
795         return result;
796 }
797 EXPORT_SYMBOL(ldlm_lock_addref_try);
798
799 /**
800  * Add specified reader/writer reference to LDLM lock \a lock.
801  * Locks LDLM lock and calls ldlm_lock_addref_internal_nolock to do the work.
802  * Only called for local locks.
803  */
804 void ldlm_lock_addref_internal(struct ldlm_lock *lock, enum ldlm_mode mode)
805 {
806         lock_res_and_lock(lock);
807         ldlm_lock_addref_internal_nolock(lock, mode);
808         unlock_res_and_lock(lock);
809 }
810
811 /**
812  * Removes reader/writer reference for LDLM lock \a lock.
813  * Assumes LDLM lock is already locked.
814  * only called in ldlm_flock_destroy and for local locks.
815  * Does NOT add lock to LRU if no r/w references left to accomodate flock locks
816  * that cannot be placed in LRU.
817  */
818 void ldlm_lock_decref_internal_nolock(struct ldlm_lock *lock,
819                                       enum ldlm_mode mode)
820 {
821         LDLM_DEBUG(lock, "ldlm_lock_decref(%s)", ldlm_lockname[mode]);
822         if (mode & (LCK_NL | LCK_CR | LCK_PR)) {
823                 LASSERT(lock->l_readers > 0);
824                 lu_ref_del(&lock->l_reference, "reader", lock);
825                 lock->l_readers--;
826         }
827         if (mode & (LCK_EX | LCK_CW | LCK_PW | LCK_GROUP | LCK_COS)) {
828                 LASSERT(lock->l_writers > 0);
829                 lu_ref_del(&lock->l_reference, "writer", lock);
830                 lock->l_writers--;
831         }
832
833         lu_ref_del(&lock->l_reference, "user", lock);
834         LDLM_LOCK_RELEASE(lock);    /* matches the LDLM_LOCK_GET() in addref */
835 }
836
837 /**
838  * Removes reader/writer reference for LDLM lock \a lock.
839  * Locks LDLM lock first.
840  * If the lock is determined to be client lock on a client and r/w refcount
841  * drops to zero and the lock is not blocked, the lock is added to LRU lock
842  * on the namespace.
843  * For blocked LDLM locks if r/w count drops to zero, blocking_ast is called.
844  */
845 void ldlm_lock_decref_internal(struct ldlm_lock *lock, enum ldlm_mode mode)
846 {
847         struct ldlm_namespace *ns;
848
849         ENTRY;
850
851         lock_res_and_lock(lock);
852
853         ns = ldlm_lock_to_ns(lock);
854
855         ldlm_lock_decref_internal_nolock(lock, mode);
856
857         if ((ldlm_is_local(lock) || lock->l_req_mode == LCK_GROUP) &&
858             !lock->l_readers && !lock->l_writers) {
859                 /* If this is a local lock on a server namespace and this was
860                  * the last reference, cancel the lock.
861                  *
862                  * Group locks are special:
863                  * They must not go in LRU, but they are not called back
864                  * like non-group locks, instead they are manually released.
865                  * They have an l_writers reference which they keep until
866                  * they are manually released, so we remove them when they have
867                  * no more reader or writer references. - LU-6368 */
868                 ldlm_set_cbpending(lock);
869         }
870
871         if (!lock->l_readers && !lock->l_writers && ldlm_is_cbpending(lock)) {
872                 unsigned int mask = D_DLMTRACE;
873
874                 /* If we received a blocked AST and this was the last reference,
875                  * run the callback. */
876                 if (ldlm_is_ns_srv(lock) && lock->l_export)
877                         mask |= D_WARNING;
878                 LDLM_DEBUG_LIMIT(mask, lock,
879                                  "final decref done on %sCBPENDING lock",
880                                  mask & D_WARNING ? "non-local " : "");
881
882                 LDLM_LOCK_GET(lock); /* dropped by bl thread */
883                 ldlm_lock_remove_from_lru(lock);
884                 unlock_res_and_lock(lock);
885
886                 if (ldlm_is_fail_loc(lock))
887                         OBD_RACE(OBD_FAIL_LDLM_CP_BL_RACE);
888
889                 if (ldlm_is_atomic_cb(lock) ||
890                     ldlm_bl_to_thread_lock(ns, NULL, lock) != 0)
891                         ldlm_handle_bl_callback(ns, NULL, lock);
892         } else if (ns_is_client(ns) &&
893                    !lock->l_readers && !lock->l_writers &&
894                    !ldlm_is_no_lru(lock) &&
895                    !ldlm_is_bl_ast(lock) &&
896                    !ldlm_is_converting(lock)) {
897
898                 /* If this is a client-side namespace and this was the last
899                  * reference, put it on the LRU.
900                  */
901                 ldlm_lock_add_to_lru(lock);
902                 unlock_res_and_lock(lock);
903                 LDLM_DEBUG(lock, "add lock into lru list");
904
905                 if (ldlm_is_fail_loc(lock))
906                         OBD_RACE(OBD_FAIL_LDLM_CP_BL_RACE);
907
908                 ldlm_pool_recalc(&ns->ns_pool, true);
909         } else {
910                 LDLM_DEBUG(lock, "do not add lock into lru list");
911                 unlock_res_and_lock(lock);
912         }
913
914         EXIT;
915 }
916
917 /**
918  * Decrease reader/writer refcount for LDLM lock with handle \a lockh
919  */
920 void ldlm_lock_decref(const struct lustre_handle *lockh, enum ldlm_mode mode)
921 {
922         struct ldlm_lock *lock = __ldlm_handle2lock(lockh, 0);
923         LASSERTF(lock != NULL, "Non-existing lock: %#llx\n", lockh->cookie);
924         ldlm_lock_decref_internal(lock, mode);
925         LDLM_LOCK_PUT(lock);
926 }
927 EXPORT_SYMBOL(ldlm_lock_decref);
928
929 /**
930  * Decrease reader/writer refcount for LDLM lock with handle
931  * \a lockh and mark it for subsequent cancellation once r/w refcount
932  * drops to zero instead of putting into LRU.
933  *
934  */
935 void ldlm_lock_decref_and_cancel(const struct lustre_handle *lockh,
936                                  enum ldlm_mode mode)
937 {
938         struct ldlm_lock *lock = __ldlm_handle2lock(lockh, 0);
939         ENTRY;
940
941         LASSERT(lock != NULL);
942
943         LDLM_DEBUG(lock, "ldlm_lock_decref(%s)", ldlm_lockname[mode]);
944         lock_res_and_lock(lock);
945         ldlm_set_cbpending(lock);
946         unlock_res_and_lock(lock);
947         ldlm_lock_decref_internal(lock, mode);
948         LDLM_LOCK_PUT(lock);
949 }
950 EXPORT_SYMBOL(ldlm_lock_decref_and_cancel);
951
952 struct sl_insert_point {
953         struct list_head *res_link;
954         struct list_head *mode_link;
955         struct list_head *policy_link;
956 };
957
958 /**
959  * Finds a position to insert the new lock into granted lock list.
960  *
961  * Used for locks eligible for skiplist optimization.
962  *
963  * Parameters:
964  *      queue [input]:  the granted list where search acts on;
965  *      req [input]:    the lock whose position to be located;
966  *      prev [output]:  positions within 3 lists to insert @req to
967  * Return Value:
968  *      filled @prev
969  * NOTE: called by
970  *  - ldlm_grant_lock_with_skiplist
971  */
972 static void search_granted_lock(struct list_head *queue,
973                                 struct ldlm_lock *req,
974                                 struct sl_insert_point *prev)
975 {
976         struct list_head *tmp;
977         struct ldlm_lock *lock, *mode_end, *policy_end;
978         ENTRY;
979
980         list_for_each(tmp, queue) {
981                 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
982
983                 mode_end = list_entry(lock->l_sl_mode.prev,
984                                           struct ldlm_lock, l_sl_mode);
985
986                 if (lock->l_req_mode != req->l_req_mode) {
987                         /* jump to last lock of mode group */
988                         tmp = &mode_end->l_res_link;
989                         continue;
990                 }
991
992                 /* suitable mode group is found */
993                 if (lock->l_resource->lr_type == LDLM_PLAIN) {
994                         /* insert point is last lock of the mode group */
995                         prev->res_link = &mode_end->l_res_link;
996                         prev->mode_link = &mode_end->l_sl_mode;
997                         prev->policy_link = &req->l_sl_policy;
998                         EXIT;
999                         return;
1000                 } else if (lock->l_resource->lr_type == LDLM_IBITS) {
1001                         for (;;) {
1002                                 policy_end =
1003                                         list_entry(lock->l_sl_policy.prev,
1004                                                        struct ldlm_lock,
1005                                                        l_sl_policy);
1006
1007                                 if (lock->l_policy_data.l_inodebits.bits ==
1008                                     req->l_policy_data.l_inodebits.bits) {
1009                                         /* insert point is last lock of
1010                                          * the policy group */
1011                                         prev->res_link =
1012                                                 &policy_end->l_res_link;
1013                                         prev->mode_link =
1014                                                 &policy_end->l_sl_mode;
1015                                         prev->policy_link =
1016                                                 &policy_end->l_sl_policy;
1017                                         EXIT;
1018                                         return;
1019                                 }
1020
1021                                 if (policy_end == mode_end)
1022                                         /* done with mode group */
1023                                         break;
1024
1025                                 /* go to next policy group within mode group */
1026                                 tmp = policy_end->l_res_link.next;
1027                                 lock = list_entry(tmp, struct ldlm_lock,
1028                                                       l_res_link);
1029                         }  /* loop over policy groups within the mode group */
1030
1031                         /* insert point is last lock of the mode group,
1032                          * new policy group is started */
1033                         prev->res_link = &mode_end->l_res_link;
1034                         prev->mode_link = &mode_end->l_sl_mode;
1035                         prev->policy_link = &req->l_sl_policy;
1036                         EXIT;
1037                         return;
1038                 } else {
1039                         LDLM_ERROR(lock,"is not LDLM_PLAIN or LDLM_IBITS lock");
1040                         LBUG();
1041                 }
1042         }
1043
1044         /* insert point is last lock on the queue,
1045          * new mode group and new policy group are started */
1046         prev->res_link = queue->prev;
1047         prev->mode_link = &req->l_sl_mode;
1048         prev->policy_link = &req->l_sl_policy;
1049         EXIT;
1050 }
1051
1052 /**
1053  * Add a lock into resource granted list after a position described by
1054  * \a prev.
1055  */
1056 static void ldlm_granted_list_add_lock(struct ldlm_lock *lock,
1057                                        struct sl_insert_point *prev)
1058 {
1059         struct ldlm_resource *res = lock->l_resource;
1060         ENTRY;
1061
1062         check_res_locked(res);
1063
1064         ldlm_resource_dump(D_INFO, res);
1065         LDLM_DEBUG(lock, "About to add lock:");
1066
1067         if (ldlm_is_destroyed(lock)) {
1068                 CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n");
1069                 return;
1070         }
1071
1072         LASSERT(list_empty(&lock->l_res_link));
1073         LASSERT(list_empty(&lock->l_sl_mode));
1074         LASSERT(list_empty(&lock->l_sl_policy));
1075
1076         /*
1077          * lock->link == prev->link means lock is first starting the group.
1078          * Don't re-add to itself to suppress kernel warnings.
1079          */
1080         if (&lock->l_res_link != prev->res_link)
1081                 list_add(&lock->l_res_link, prev->res_link);
1082         if (&lock->l_sl_mode != prev->mode_link)
1083                 list_add(&lock->l_sl_mode, prev->mode_link);
1084         if (&lock->l_sl_policy != prev->policy_link)
1085                 list_add(&lock->l_sl_policy, prev->policy_link);
1086
1087         EXIT;
1088 }
1089
1090 /**
1091  * Add a lock to granted list on a resource maintaining skiplist
1092  * correctness.
1093  */
1094 void ldlm_grant_lock_with_skiplist(struct ldlm_lock *lock)
1095 {
1096         struct sl_insert_point prev;
1097
1098         LASSERT(ldlm_is_granted(lock));
1099
1100         search_granted_lock(&lock->l_resource->lr_granted, lock, &prev);
1101         ldlm_granted_list_add_lock(lock, &prev);
1102 }
1103
1104 /**
1105  * Perform lock granting bookkeeping.
1106  *
1107  * Includes putting the lock into granted list and updating lock mode.
1108  * NOTE: called by
1109  *  - ldlm_lock_enqueue
1110  *  - ldlm_reprocess_queue
1111  *
1112  * must be called with lr_lock held
1113  */
1114 void ldlm_grant_lock(struct ldlm_lock *lock, struct list_head *work_list)
1115 {
1116         struct ldlm_resource *res = lock->l_resource;
1117         ENTRY;
1118
1119         check_res_locked(res);
1120
1121         lock->l_granted_mode = lock->l_req_mode;
1122
1123         if (work_list && lock->l_completion_ast != NULL)
1124                 ldlm_add_ast_work_item(lock, NULL, work_list);
1125
1126         if (res->lr_type == LDLM_PLAIN || res->lr_type == LDLM_IBITS)
1127                 ldlm_grant_lock_with_skiplist(lock);
1128         else if (res->lr_type == LDLM_EXTENT)
1129                 ldlm_extent_add_lock(res, lock);
1130         else if (res->lr_type == LDLM_FLOCK) {
1131                 /* We should not add locks to granted list in the following
1132                  * cases:
1133                  * - this is an UNLOCK but not a real lock;
1134                  * - this is a TEST lock;
1135                  * - this is a F_CANCELLK lock (async flock has req_mode == 0)
1136                  * - this is a deadlock (flock cannot be granted) */
1137                 if (lock->l_req_mode == 0 ||
1138                     lock->l_req_mode == LCK_NL ||
1139                     ldlm_is_test_lock(lock) ||
1140                     ldlm_is_flock_deadlock(lock))
1141                         RETURN_EXIT;
1142                 ldlm_resource_add_lock(res, &res->lr_granted, lock);
1143         } else {
1144                 LBUG();
1145         }
1146
1147         ldlm_pool_add(&ldlm_res_to_ns(res)->ns_pool, lock);
1148         EXIT;
1149 }
1150
1151 /**
1152  * Check if the given @lock meets the criteria for a match.
1153  * A reference on the lock is taken if matched.
1154  *
1155  * @lock        test-against this lock
1156  * @data        parameters
1157  *
1158  * RETURN       returns true if @lock matches @data, false otherwise
1159  */
1160 static bool lock_matches(struct ldlm_lock *lock, struct ldlm_match_data *data)
1161 {
1162         union ldlm_policy_data *lpol = &lock->l_policy_data;
1163         enum ldlm_mode match = LCK_MINMODE;
1164
1165         if (lock == data->lmd_old)
1166                 return true;
1167
1168         /* Check if this lock can be matched.
1169          * Used by LU-2919(exclusive open) for open lease lock */
1170         if (ldlm_is_excl(lock))
1171                 return false;
1172
1173         /* llite sometimes wants to match locks that will be
1174          * canceled when their users drop, but we allow it to match
1175          * if it passes in CBPENDING and the lock still has users.
1176          * this is generally only going to be used by children
1177          * whose parents already hold a lock so forward progress
1178          * can still happen. */
1179         if (ldlm_is_cbpending(lock) &&
1180             !(data->lmd_flags & LDLM_FL_CBPENDING))
1181                 return false;
1182
1183         if (!(data->lmd_match & LDLM_MATCH_UNREF) && ldlm_is_cbpending(lock) &&
1184             lock->l_readers == 0 && lock->l_writers == 0)
1185                 return false;
1186
1187         if (!(lock->l_req_mode & *data->lmd_mode))
1188                 return false;
1189
1190         /* When we search for ast_data, we are not doing a traditional match,
1191          * so we don't worry about IBITS or extent matching.
1192          */
1193         if (data->lmd_match & (LDLM_MATCH_AST | LDLM_MATCH_AST_ANY)) {
1194                 if (!lock->l_ast_data)
1195                         return false;
1196
1197                 if (data->lmd_match & LDLM_MATCH_AST_ANY)
1198                         goto matched;
1199         }
1200
1201         match = lock->l_req_mode;
1202
1203         switch (lock->l_resource->lr_type) {
1204         case LDLM_EXTENT:
1205                 if (!(data->lmd_match & LDLM_MATCH_RIGHT) &&
1206                     (lpol->l_extent.start > data->lmd_policy->l_extent.start ||
1207                      lpol->l_extent.end < data->lmd_policy->l_extent.end))
1208                         return false;
1209
1210                 if (unlikely(match == LCK_GROUP) &&
1211                     data->lmd_policy->l_extent.gid != LDLM_GID_ANY &&
1212                     lpol->l_extent.gid != data->lmd_policy->l_extent.gid)
1213                         return false;
1214                 break;
1215         case LDLM_IBITS:
1216                 /* We match if we have existing lock with same or wider set
1217                    of bits. */
1218                 if ((lpol->l_inodebits.bits &
1219                      data->lmd_policy->l_inodebits.bits) !=
1220                     data->lmd_policy->l_inodebits.bits)
1221                         return false;
1222
1223                 if (unlikely(match == LCK_GROUP) &&
1224                     data->lmd_policy->l_inodebits.li_gid != LDLM_GID_ANY &&
1225                     lpol->l_inodebits.li_gid !=
1226                     data->lmd_policy->l_inodebits.li_gid)
1227                         return false;
1228                 break;
1229         default:
1230                 ;
1231         }
1232
1233         /* We match if we have existing lock with same or wider set
1234            of bits. */
1235         if (!(data->lmd_match & LDLM_MATCH_UNREF) && LDLM_HAVE_MASK(lock, GONE))
1236                 return false;
1237
1238         if (!equi(data->lmd_flags & LDLM_FL_LOCAL_ONLY, ldlm_is_local(lock)))
1239                 return false;
1240
1241         /* Filter locks by skipping flags */
1242         if (data->lmd_skip_flags & lock->l_flags)
1243                 return false;
1244
1245 matched:
1246         if (data->lmd_flags & LDLM_FL_TEST_LOCK) {
1247                 LDLM_LOCK_GET(lock);
1248                 ldlm_lock_touch_in_lru(lock);
1249         } else {
1250                 ldlm_lock_addref_internal_nolock(lock, match);
1251         }
1252
1253         *data->lmd_mode = match;
1254         data->lmd_lock = lock;
1255
1256         return true;
1257 }
1258
1259 static unsigned int itree_overlap_cb(struct interval_node *in, void *args)
1260 {
1261         struct ldlm_interval *node = to_ldlm_interval(in);
1262         struct ldlm_match_data *data = args;
1263         struct ldlm_lock *lock;
1264
1265         list_for_each_entry(lock, &node->li_group, l_sl_policy) {
1266                 if (lock_matches(lock, data))
1267                         return INTERVAL_ITER_STOP;
1268         }
1269         return INTERVAL_ITER_CONT;
1270 }
1271
1272 /**
1273  * Search for a lock with given parameters in interval trees.
1274  *
1275  * \param res      search for a lock in this resource
1276  * \param data     parameters
1277  *
1278  * \retval a referenced lock or NULL.
1279  */
1280 struct ldlm_lock *search_itree(struct ldlm_resource *res,
1281                                struct ldlm_match_data *data)
1282 {
1283         struct interval_node_extent ext = {
1284                 .start     = data->lmd_policy->l_extent.start,
1285                 .end       = data->lmd_policy->l_extent.end
1286         };
1287         int idx;
1288
1289         data->lmd_lock = NULL;
1290
1291         if (data->lmd_match & LDLM_MATCH_RIGHT)
1292                 ext.end = OBD_OBJECT_EOF;
1293
1294         for (idx = 0; idx < LCK_MODE_NUM; idx++) {
1295                 struct ldlm_interval_tree *tree = &res->lr_itree[idx];
1296
1297                 if (tree->lit_root == NULL)
1298                         continue;
1299
1300                 if (!(tree->lit_mode & *data->lmd_mode))
1301                         continue;
1302
1303                 interval_search(tree->lit_root, &ext,
1304                                 itree_overlap_cb, data);
1305                 if (data->lmd_lock)
1306                         return data->lmd_lock;
1307         }
1308
1309         return NULL;
1310 }
1311 EXPORT_SYMBOL(search_itree);
1312
1313
1314 /**
1315  * Search for a lock with given properties in a queue.
1316  *
1317  * \param queue    search for a lock in this queue
1318  * \param data     parameters
1319  *
1320  * \retval a referenced lock or NULL.
1321  */
1322 static struct ldlm_lock *search_queue(struct list_head *queue,
1323                                       struct ldlm_match_data *data)
1324 {
1325         struct ldlm_lock *lock;
1326
1327         data->lmd_lock = NULL;
1328
1329         list_for_each_entry(lock, queue, l_res_link)
1330                 if (lock_matches(lock, data))
1331                         return data->lmd_lock;
1332
1333         return NULL;
1334 }
1335
1336 void ldlm_lock_fail_match_locked(struct ldlm_lock *lock)
1337 {
1338         if ((lock->l_flags & LDLM_FL_FAIL_NOTIFIED) == 0) {
1339                 lock->l_flags |= LDLM_FL_FAIL_NOTIFIED;
1340                 wake_up_all(&lock->l_waitq);
1341         }
1342 }
1343 EXPORT_SYMBOL(ldlm_lock_fail_match_locked);
1344
1345 void ldlm_lock_fail_match(struct ldlm_lock *lock)
1346 {
1347         lock_res_and_lock(lock);
1348         ldlm_lock_fail_match_locked(lock);
1349         unlock_res_and_lock(lock);
1350 }
1351
1352 /**
1353  * Mark lock as "matchable" by OST.
1354  *
1355  * Used to prevent certain races in LOV/OSC where the lock is granted, but LVB
1356  * is not yet valid.
1357  * Assumes LDLM lock is already locked.
1358  */
1359 void ldlm_lock_allow_match_locked(struct ldlm_lock *lock)
1360 {
1361         ldlm_set_lvb_ready(lock);
1362         wake_up_all(&lock->l_waitq);
1363 }
1364 EXPORT_SYMBOL(ldlm_lock_allow_match_locked);
1365
1366 /**
1367  * Mark lock as "matchable" by OST.
1368  * Locks the lock and then \see ldlm_lock_allow_match_locked
1369  */
1370 void ldlm_lock_allow_match(struct ldlm_lock *lock)
1371 {
1372         lock_res_and_lock(lock);
1373         ldlm_lock_allow_match_locked(lock);
1374         unlock_res_and_lock(lock);
1375 }
1376 EXPORT_SYMBOL(ldlm_lock_allow_match);
1377
1378 /**
1379  * Attempt to find a lock with specified properties.
1380  *
1381  * Typically returns a reference to matched lock unless LDLM_FL_TEST_LOCK is
1382  * set in \a flags
1383  *
1384  * Can be called in two ways:
1385  *
1386  * If 'ns' is NULL, then lockh describes an existing lock that we want to look
1387  * for a duplicate of.
1388  *
1389  * Otherwise, all of the fields must be filled in, to match against.
1390  *
1391  * If 'flags' contains LDLM_FL_LOCAL_ONLY, then only match local locks on the
1392  *     server (ie, connh is NULL)
1393  * If 'flags' contains LDLM_FL_BLOCK_GRANTED, then only locks on the granted
1394  *     list will be considered
1395  * If 'flags' contains LDLM_FL_CBPENDING, then locks that have been marked
1396  *     to be canceled can still be matched as long as they still have reader
1397  *     or writer refernces
1398  * If 'flags' contains LDLM_FL_TEST_LOCK, then don't actually reference a lock,
1399  *     just tell us if we would have matched.
1400  *
1401  * \retval 1 if it finds an already-existing lock that is compatible; in this
1402  * case, lockh is filled in with a addref()ed lock
1403  *
1404  * We also check security context, and if that fails we simply return 0 (to
1405  * keep caller code unchanged), the context failure will be discovered by
1406  * caller sometime later.
1407  */
1408 enum ldlm_mode ldlm_lock_match_with_skip(struct ldlm_namespace *ns,
1409                                          __u64 flags, __u64 skip_flags,
1410                                          const struct ldlm_res_id *res_id,
1411                                          enum ldlm_type type,
1412                                          union ldlm_policy_data *policy,
1413                                          enum ldlm_mode mode,
1414                                          struct lustre_handle *lockh,
1415                                          enum ldlm_match_flags match_flags)
1416 {
1417         struct ldlm_match_data data = {
1418                 .lmd_old = NULL,
1419                 .lmd_lock = NULL,
1420                 .lmd_mode = &mode,
1421                 .lmd_policy = policy,
1422                 .lmd_flags = flags,
1423                 .lmd_skip_flags = skip_flags,
1424                 .lmd_match = match_flags,
1425         };
1426         struct ldlm_resource *res;
1427         struct ldlm_lock *lock;
1428         int matched;
1429
1430         ENTRY;
1431
1432         if (ns == NULL) {
1433                 data.lmd_old = ldlm_handle2lock(lockh);
1434                 LASSERT(data.lmd_old != NULL);
1435
1436                 ns = ldlm_lock_to_ns(data.lmd_old);
1437                 res_id = &data.lmd_old->l_resource->lr_name;
1438                 type = data.lmd_old->l_resource->lr_type;
1439                 *data.lmd_mode = data.lmd_old->l_req_mode;
1440         }
1441
1442         res = ldlm_resource_get(ns, NULL, res_id, type, 0);
1443         if (IS_ERR(res)) {
1444                 LASSERT(data.lmd_old == NULL);
1445                 RETURN(0);
1446         }
1447
1448         LDLM_RESOURCE_ADDREF(res);
1449         lock_res(res);
1450         if (res->lr_type == LDLM_EXTENT)
1451                 lock = search_itree(res, &data);
1452         else
1453                 lock = search_queue(&res->lr_granted, &data);
1454         if (!lock && !(flags & LDLM_FL_BLOCK_GRANTED))
1455                 lock = search_queue(&res->lr_waiting, &data);
1456         matched = lock ? mode : 0;
1457         unlock_res(res);
1458         LDLM_RESOURCE_DELREF(res);
1459         ldlm_resource_putref(res);
1460
1461         if (lock) {
1462                 ldlm_lock2handle(lock, lockh);
1463                 if ((flags & LDLM_FL_LVB_READY) &&
1464                     (!ldlm_is_lvb_ready(lock))) {
1465                         __u64 wait_flags = LDLM_FL_LVB_READY |
1466                                 LDLM_FL_DESTROYED | LDLM_FL_FAIL_NOTIFIED;
1467
1468                         if (lock->l_completion_ast) {
1469                                 int err = lock->l_completion_ast(lock,
1470                                                         LDLM_FL_WAIT_NOREPROC,
1471                                                         NULL);
1472                                 if (err)
1473                                         GOTO(out_fail_match, matched = 0);
1474                         }
1475
1476                         wait_event_idle_timeout(
1477                                 lock->l_waitq,
1478                                 lock->l_flags & wait_flags,
1479                                 cfs_time_seconds(obd_timeout));
1480
1481                         if (!ldlm_is_lvb_ready(lock))
1482                                 GOTO(out_fail_match, matched = 0);
1483                 }
1484
1485                 /* check user's security context */
1486                 if (lock->l_conn_export &&
1487                     sptlrpc_import_check_ctx(
1488                                 class_exp2cliimp(lock->l_conn_export)))
1489                         GOTO(out_fail_match, matched = 0);
1490
1491                 LDLM_DEBUG(lock, "matched (%llu %llu)",
1492                            (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1493                            res_id->name[2] : policy->l_extent.start,
1494                            (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1495                            res_id->name[3] : policy->l_extent.end);
1496
1497 out_fail_match:
1498                 if (flags & LDLM_FL_TEST_LOCK)
1499                         LDLM_LOCK_RELEASE(lock);
1500                 else if (!matched)
1501                         ldlm_lock_decref_internal(lock, mode);
1502         }
1503
1504         /* less verbose for test-only */
1505         if (!matched && !(flags & LDLM_FL_TEST_LOCK)) {
1506                 LDLM_DEBUG_NOLOCK("not matched ns %p type %u mode %u res "
1507                                   "%llu/%llu (%llu %llu)", ns,
1508                                   type, mode, res_id->name[0], res_id->name[1],
1509                                   (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1510                                   res_id->name[2] : policy->l_extent.start,
1511                                   (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1512                                   res_id->name[3] : policy->l_extent.end);
1513         }
1514         if (data.lmd_old != NULL)
1515                 LDLM_LOCK_PUT(data.lmd_old);
1516
1517         return matched;
1518 }
1519 EXPORT_SYMBOL(ldlm_lock_match_with_skip);
1520
1521 enum ldlm_mode ldlm_revalidate_lock_handle(const struct lustre_handle *lockh,
1522                                            __u64 *bits)
1523 {
1524         struct ldlm_lock *lock;
1525         enum ldlm_mode mode = 0;
1526         ENTRY;
1527
1528         lock = ldlm_handle2lock(lockh);
1529         if (lock != NULL) {
1530                 lock_res_and_lock(lock);
1531                 if (LDLM_HAVE_MASK(lock, GONE))
1532                         GOTO(out, mode);
1533
1534                 if (ldlm_is_cbpending(lock) &&
1535                     lock->l_readers == 0 && lock->l_writers == 0)
1536                         GOTO(out, mode);
1537
1538                 if (bits)
1539                         *bits = lock->l_policy_data.l_inodebits.bits;
1540                 mode = lock->l_granted_mode;
1541                 ldlm_lock_addref_internal_nolock(lock, mode);
1542         }
1543
1544         EXIT;
1545
1546 out:
1547         if (lock != NULL) {
1548                 unlock_res_and_lock(lock);
1549                 LDLM_LOCK_PUT(lock);
1550         }
1551         return mode;
1552 }
1553 EXPORT_SYMBOL(ldlm_revalidate_lock_handle);
1554
1555 /** The caller must guarantee that the buffer is large enough. */
1556 int ldlm_fill_lvb(struct ldlm_lock *lock, struct req_capsule *pill,
1557                   enum req_location loc, void *data, int size)
1558 {
1559         void *lvb;
1560         ENTRY;
1561
1562         LASSERT(data != NULL);
1563         LASSERT(size >= 0);
1564
1565         switch (lock->l_lvb_type) {
1566         case LVB_T_OST:
1567                 if (size == sizeof(struct ost_lvb)) {
1568                         if (loc == RCL_CLIENT)
1569                                 lvb = req_capsule_client_swab_get(pill,
1570                                                 &RMF_DLM_LVB,
1571                                                 lustre_swab_ost_lvb);
1572                         else
1573                                 lvb = req_capsule_server_swab_get(pill,
1574                                                 &RMF_DLM_LVB,
1575                                                 lustre_swab_ost_lvb);
1576                         if (unlikely(lvb == NULL)) {
1577                                 LDLM_ERROR(lock, "no LVB");
1578                                 RETURN(-EPROTO);
1579                         }
1580
1581                         memcpy(data, lvb, size);
1582                 } else if (size == sizeof(struct ost_lvb_v1)) {
1583                         struct ost_lvb *olvb = data;
1584
1585                         if (loc == RCL_CLIENT)
1586                                 lvb = req_capsule_client_swab_get(pill,
1587                                                 &RMF_DLM_LVB,
1588                                                 lustre_swab_ost_lvb_v1);
1589                         else
1590                                 lvb = req_capsule_server_sized_swab_get(pill,
1591                                                 &RMF_DLM_LVB, size,
1592                                                 lustre_swab_ost_lvb_v1);
1593                         if (unlikely(lvb == NULL)) {
1594                                 LDLM_ERROR(lock, "no LVB");
1595                                 RETURN(-EPROTO);
1596                         }
1597
1598                         memcpy(data, lvb, size);
1599                         olvb->lvb_mtime_ns = 0;
1600                         olvb->lvb_atime_ns = 0;
1601                         olvb->lvb_ctime_ns = 0;
1602                 } else {
1603                         LDLM_ERROR(lock, "Replied unexpected ost LVB size %d",
1604                                    size);
1605                         RETURN(-EINVAL);
1606                 }
1607                 break;
1608         case LVB_T_LQUOTA:
1609                 if (size == sizeof(struct lquota_lvb)) {
1610                         if (loc == RCL_CLIENT)
1611                                 lvb = req_capsule_client_swab_get(pill,
1612                                                 &RMF_DLM_LVB,
1613                                                 lustre_swab_lquota_lvb);
1614                         else
1615                                 lvb = req_capsule_server_swab_get(pill,
1616                                                 &RMF_DLM_LVB,
1617                                                 lustre_swab_lquota_lvb);
1618                         if (unlikely(lvb == NULL)) {
1619                                 LDLM_ERROR(lock, "no LVB");
1620                                 RETURN(-EPROTO);
1621                         }
1622
1623                         memcpy(data, lvb, size);
1624                 } else {
1625                         LDLM_ERROR(lock, "Replied unexpected lquota LVB size %d",
1626                                    size);
1627                         RETURN(-EINVAL);
1628                 }
1629                 break;
1630         case LVB_T_LAYOUT:
1631                 if (size == 0)
1632                         break;
1633
1634                 if (loc == RCL_CLIENT)
1635                         lvb = req_capsule_client_get(pill, &RMF_DLM_LVB);
1636                 else
1637                         lvb = req_capsule_server_get(pill, &RMF_DLM_LVB);
1638                 if (unlikely(lvb == NULL)) {
1639                         LDLM_ERROR(lock, "no LVB");
1640                         RETURN(-EPROTO);
1641                 }
1642
1643                 memcpy(data, lvb, size);
1644                 break;
1645         default:
1646                 LDLM_ERROR(lock, "Unknown LVB type: %d", lock->l_lvb_type);
1647                 libcfs_debug_dumpstack(NULL);
1648                 RETURN(-EINVAL);
1649         }
1650
1651         RETURN(0);
1652 }
1653
1654 /**
1655  * Create and fill in new LDLM lock with specified properties.
1656  * Returns a referenced lock
1657  */
1658 struct ldlm_lock *ldlm_lock_create(struct ldlm_namespace *ns,
1659                                    const struct ldlm_res_id *res_id,
1660                                    enum ldlm_type type,
1661                                    enum ldlm_mode mode,
1662                                    const struct ldlm_callback_suite *cbs,
1663                                    void *data, __u32 lvb_len,
1664                                    enum lvb_type lvb_type)
1665 {
1666         struct ldlm_lock        *lock;
1667         struct ldlm_resource    *res;
1668         int                     rc;
1669         ENTRY;
1670
1671         res = ldlm_resource_get(ns, NULL, res_id, type, 1);
1672         if (IS_ERR(res))
1673                 RETURN(ERR_CAST(res));
1674
1675         lock = ldlm_lock_new(res);
1676         if (lock == NULL)
1677                 RETURN(ERR_PTR(-ENOMEM));
1678
1679         lock->l_req_mode = mode;
1680         lock->l_ast_data = data;
1681         lock->l_pid = current->pid;
1682         if (ns_is_server(ns))
1683                 ldlm_set_ns_srv(lock);
1684         if (cbs) {
1685                 lock->l_blocking_ast = cbs->lcs_blocking;
1686                 lock->l_completion_ast = cbs->lcs_completion;
1687                 lock->l_glimpse_ast = cbs->lcs_glimpse;
1688         }
1689
1690         switch (type) {
1691         case LDLM_EXTENT:
1692                 rc = ldlm_extent_alloc_lock(lock);
1693                 break;
1694         case LDLM_IBITS:
1695                 rc = ldlm_inodebits_alloc_lock(lock);
1696                 break;
1697         default:
1698                 rc = 0;
1699         }
1700         if (rc)
1701                 GOTO(out, rc);
1702
1703         if (lvb_len) {
1704                 lock->l_lvb_len = lvb_len;
1705                 OBD_ALLOC_LARGE(lock->l_lvb_data, lvb_len);
1706                 if (lock->l_lvb_data == NULL)
1707                         GOTO(out, rc = -ENOMEM);
1708         }
1709
1710         lock->l_lvb_type = lvb_type;
1711         if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_NEW_LOCK))
1712                 GOTO(out, rc = -ENOENT);
1713
1714         RETURN(lock);
1715
1716 out:
1717         ldlm_lock_destroy(lock);
1718         LDLM_LOCK_RELEASE(lock);
1719         RETURN(ERR_PTR(rc));
1720 }
1721
1722 #ifdef HAVE_SERVER_SUPPORT
1723 static enum ldlm_error ldlm_lock_enqueue_helper(struct ldlm_lock *lock,
1724                                              __u64 *flags)
1725 {
1726         struct ldlm_resource *res = lock->l_resource;
1727         enum ldlm_error rc = ELDLM_OK;
1728         LIST_HEAD(rpc_list);
1729         ldlm_processing_policy policy;
1730
1731         ENTRY;
1732
1733         policy = ldlm_get_processing_policy(res);
1734         policy(lock, flags, LDLM_PROCESS_ENQUEUE, &rc, &rpc_list);
1735         if (rc == ELDLM_OK && lock->l_granted_mode != lock->l_req_mode &&
1736             res->lr_type != LDLM_FLOCK)
1737                 rc = ldlm_handle_conflict_lock(lock, flags, &rpc_list);
1738
1739         if (!list_empty(&rpc_list))
1740                 ldlm_discard_bl_list(&rpc_list);
1741
1742         RETURN(rc);
1743 }
1744 #endif
1745
1746 /**
1747  * Enqueue (request) a lock.
1748  *
1749  * Does not block. As a result of enqueue the lock would be put
1750  * into granted or waiting list.
1751  *
1752  * If namespace has intent policy sent and the lock has LDLM_FL_HAS_INTENT flag
1753  * set, skip all the enqueueing and delegate lock processing to intent policy
1754  * function.
1755  */
1756 enum ldlm_error ldlm_lock_enqueue(const struct lu_env *env,
1757                                   struct ldlm_namespace *ns,
1758                                   struct ldlm_lock **lockp,
1759                                   void *cookie, __u64 *flags)
1760 {
1761         struct ldlm_lock *lock = *lockp;
1762         struct ldlm_resource *res;
1763         int local = ns_is_client(ns);
1764         enum ldlm_error rc = ELDLM_OK;
1765         struct ldlm_interval *node = NULL;
1766 #ifdef HAVE_SERVER_SUPPORT
1767         bool reconstruct = false;
1768 #endif
1769         ENTRY;
1770
1771         /* policies are not executed on the client or during replay */
1772         if ((*flags & (LDLM_FL_HAS_INTENT|LDLM_FL_REPLAY)) == LDLM_FL_HAS_INTENT
1773             && !local && ns->ns_policy) {
1774                 rc = ns->ns_policy(env, ns, lockp, cookie, lock->l_req_mode,
1775                                    *flags, NULL);
1776                 if (rc == ELDLM_LOCK_REPLACED) {
1777                         /* The lock that was returned has already been granted,
1778                          * and placed into lockp.  If it's not the same as the
1779                          * one we passed in, then destroy the old one and our
1780                          * work here is done. */
1781                         if (lock != *lockp) {
1782                                 ldlm_lock_destroy(lock);
1783                                 LDLM_LOCK_RELEASE(lock);
1784                         }
1785                         *flags |= LDLM_FL_LOCK_CHANGED;
1786                         RETURN(0);
1787                 } else if (rc != ELDLM_OK &&
1788                            ldlm_is_granted(lock)) {
1789                         LASSERT(*flags & LDLM_FL_RESENT);
1790                         /* It may happen that ns_policy returns an error in
1791                          * resend case, object may be unlinked or just some
1792                          * error occurs. It is unclear if lock reached the
1793                          * client in the original reply, just leave the lock on
1794                          * server, not returning it again to client. Due to
1795                          * LU-6529, the server will not OOM. */
1796                         RETURN(rc);
1797                 } else if (rc != ELDLM_OK ||
1798                            (rc == ELDLM_OK && (*flags & LDLM_FL_INTENT_ONLY))) {
1799                         ldlm_lock_destroy(lock);
1800                         RETURN(rc);
1801                 }
1802         }
1803
1804         if (*flags & LDLM_FL_RESENT) {
1805                 /* Reconstruct LDLM_FL_SRV_ENQ_MASK @flags for reply.
1806                  * Set LOCK_CHANGED always.
1807                  * Check if the lock is granted for BLOCK_GRANTED.
1808                  * Take NO_TIMEOUT from the lock as it is inherited through
1809                  * LDLM_FL_INHERIT_MASK */
1810                 *flags |= LDLM_FL_LOCK_CHANGED;
1811                 if (!ldlm_is_granted(lock))
1812                         *flags |= LDLM_FL_BLOCK_GRANTED;
1813                 *flags |= lock->l_flags & LDLM_FL_NO_TIMEOUT;
1814                 RETURN(ELDLM_OK);
1815         }
1816
1817 #ifdef HAVE_SERVER_SUPPORT
1818         /* For a replaying lock, it might be already in granted list. So
1819          * unlinking the lock will cause the interval node to be freed, we
1820          * have to allocate the interval node early otherwise we can't regrant
1821          * this lock in the future. - jay
1822          *
1823          * The only time the ldlm_resource changes for the ldlm_lock is when
1824          * ldlm_lock_change_resource() is called and that only happens for
1825          * the Lustre client case.
1826          */
1827         if (!local && (*flags & LDLM_FL_REPLAY) &&
1828             lock->l_resource->lr_type == LDLM_EXTENT)
1829                 OBD_SLAB_ALLOC_PTR_GFP(node, ldlm_interval_slab, GFP_NOFS);
1830
1831         reconstruct = !local && lock->l_resource->lr_type == LDLM_FLOCK &&
1832                       !(*flags & LDLM_FL_TEST_LOCK);
1833         if (reconstruct) {
1834                 rc = req_can_reconstruct(cookie, NULL);
1835                 if (rc != 0) {
1836                         if (rc == 1)
1837                                 rc = 0;
1838                         RETURN(rc);
1839                 }
1840         }
1841 #endif
1842         res = lock_res_and_lock(lock);
1843         if (local && ldlm_is_granted(lock)) {
1844                 /* The server returned a blocked lock, but it was granted
1845                  * before we got a chance to actually enqueue it.  We don't
1846                  * need to do anything else. */
1847                 *flags &= ~LDLM_FL_BLOCKED_MASK;
1848                 GOTO(out, rc = ELDLM_OK);
1849         }
1850
1851         ldlm_resource_unlink_lock(lock);
1852         if (res->lr_type == LDLM_EXTENT && lock->l_tree_node == NULL) {
1853                 if (node == NULL) {
1854                         ldlm_lock_destroy_nolock(lock);
1855                         GOTO(out, rc = -ENOMEM);
1856                 }
1857
1858                 INIT_LIST_HEAD(&node->li_group);
1859                 ldlm_interval_attach(node, lock);
1860                 node = NULL;
1861         }
1862
1863         /* Some flags from the enqueue want to make it into the AST, via the
1864          * lock's l_flags. */
1865         if (*flags & LDLM_FL_AST_DISCARD_DATA)
1866                 ldlm_set_ast_discard_data(lock);
1867         if (*flags & LDLM_FL_TEST_LOCK)
1868                 ldlm_set_test_lock(lock);
1869         if (*flags & LDLM_FL_COS_INCOMPAT)
1870                 ldlm_set_cos_incompat(lock);
1871         if (*flags & LDLM_FL_COS_ENABLED)
1872                 ldlm_set_cos_enabled(lock);
1873
1874         /* This distinction between local lock trees is very important; a client
1875          * namespace only has information about locks taken by that client, and
1876          * thus doesn't have enough information to decide for itself if it can
1877          * be granted (below).  In this case, we do exactly what the server
1878          * tells us to do, as dictated by the 'flags'.
1879          *
1880          * We do exactly the same thing during recovery, when the server is
1881          * more or less trusting the clients not to lie.
1882          *
1883          * FIXME (bug 268): Detect obvious lies by checking compatibility in
1884          * granted queue. */
1885         if (local) {
1886                 if (*flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED))
1887                         ldlm_resource_add_lock(res, &res->lr_waiting, lock);
1888                 else
1889                         ldlm_grant_lock(lock, NULL);
1890                 GOTO(out, rc = ELDLM_OK);
1891 #ifdef HAVE_SERVER_SUPPORT
1892         } else if (*flags & LDLM_FL_REPLAY) {
1893                 if (*flags & LDLM_FL_BLOCK_WAIT) {
1894                         ldlm_resource_add_lock(res, &res->lr_waiting, lock);
1895                         GOTO(out, rc = ELDLM_OK);
1896                 } else if (*flags & LDLM_FL_BLOCK_GRANTED) {
1897                         ldlm_grant_lock(lock, NULL);
1898                         GOTO(out, rc = ELDLM_OK);
1899                 }
1900                 /* If no flags, fall through to normal enqueue path. */
1901         }
1902
1903         rc = ldlm_lock_enqueue_helper(lock, flags);
1904         GOTO(out, rc);
1905 #else
1906         } else {
1907                 CERROR("This is client-side-only module, cannot handle "
1908                        "LDLM_NAMESPACE_SERVER resource type lock.\n");
1909                 LBUG();
1910         }
1911 #endif
1912
1913 out:
1914         unlock_res_and_lock(lock);
1915
1916 #ifdef HAVE_SERVER_SUPPORT
1917         if (reconstruct) {
1918                 struct ptlrpc_request *req = cookie;
1919
1920                 tgt_mk_reply_data(NULL, NULL,
1921                                   &req->rq_export->exp_target_data,
1922                                   req, 0, NULL, false, 0);
1923         }
1924 #endif
1925         if (node)
1926                 OBD_SLAB_FREE(node, ldlm_interval_slab, sizeof(*node));
1927         return rc;
1928 }
1929
1930 #ifdef HAVE_SERVER_SUPPORT
1931 /**
1932  * Iterate through all waiting locks on a given resource queue and attempt to
1933  * grant them.
1934  *
1935  * Must be called with resource lock held.
1936  */
1937 int ldlm_reprocess_queue(struct ldlm_resource *res, struct list_head *queue,
1938                          struct list_head *work_list,
1939                          enum ldlm_process_intention intention,
1940                          struct ldlm_lock *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, NULL);
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                                  struct ldlm_lock *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                 goto restart;
2387         }
2388 #else
2389         ENTRY;
2390
2391         if (!ns_is_client(ldlm_res_to_ns(res))) {
2392                 CERROR("This is client-side-only module, cannot handle "
2393                        "LDLM_NAMESPACE_SERVER resource type lock.\n");
2394                 LBUG();
2395         }
2396 #endif
2397         EXIT;
2398 }
2399
2400 void ldlm_reprocess_all(struct ldlm_resource *res, struct ldlm_lock *hint)
2401 {
2402         __ldlm_reprocess_all(res, LDLM_PROCESS_RESCAN, hint);
2403 }
2404 EXPORT_SYMBOL(ldlm_reprocess_all);
2405
2406 static int ldlm_reprocess_res(struct cfs_hash *hs, struct cfs_hash_bd *bd,
2407                               struct hlist_node *hnode, void *arg)
2408 {
2409         struct ldlm_resource *res = cfs_hash_object(hs, hnode);
2410
2411         /* This is only called once after recovery done. LU-8306. */
2412         __ldlm_reprocess_all(res, LDLM_PROCESS_RECOVERY, NULL);
2413         return 0;
2414 }
2415
2416 /**
2417  * Iterate through all resources on a namespace attempting to grant waiting
2418  * locks.
2419  */
2420 void ldlm_reprocess_recovery_done(struct ldlm_namespace *ns)
2421 {
2422         ENTRY;
2423
2424         if (ns != NULL) {
2425                 cfs_hash_for_each_nolock(ns->ns_rs_hash,
2426                                          ldlm_reprocess_res, NULL, 0);
2427         }
2428         EXIT;
2429 }
2430
2431 /**
2432  * Helper function to call blocking AST for LDLM lock \a lock in a
2433  * "cancelling" mode.
2434  */
2435 void ldlm_cancel_callback(struct ldlm_lock *lock)
2436 {
2437         check_res_locked(lock->l_resource);
2438         if (!ldlm_is_cancel(lock)) {
2439                 ldlm_set_cancel(lock);
2440                 if (lock->l_blocking_ast) {
2441                         unlock_res_and_lock(lock);
2442                         lock->l_blocking_ast(lock, NULL, lock->l_ast_data,
2443                                              LDLM_CB_CANCELING);
2444                         lock_res_and_lock(lock);
2445                 } else {
2446                         LDLM_DEBUG(lock, "no blocking ast");
2447                 }
2448
2449                 /* only canceller can set bl_done bit */
2450                 ldlm_set_bl_done(lock);
2451                 wake_up_all(&lock->l_waitq);
2452         } else if (!ldlm_is_bl_done(lock)) {
2453                 /* The lock is guaranteed to have been canceled once
2454                  * returning from this function. */
2455                 unlock_res_and_lock(lock);
2456                 wait_event_idle(lock->l_waitq, is_bl_done(lock));
2457                 lock_res_and_lock(lock);
2458         }
2459 }
2460
2461 /**
2462  * Remove skiplist-enabled LDLM lock \a req from granted list
2463  */
2464 void ldlm_unlink_lock_skiplist(struct ldlm_lock *req)
2465 {
2466         if (req->l_resource->lr_type != LDLM_PLAIN &&
2467             req->l_resource->lr_type != LDLM_IBITS)
2468                 return;
2469
2470         list_del_init(&req->l_sl_policy);
2471         list_del_init(&req->l_sl_mode);
2472 }
2473
2474 /**
2475  * Attempts to cancel LDLM lock \a lock that has no reader/writer references.
2476  */
2477 void ldlm_lock_cancel(struct ldlm_lock *lock)
2478 {
2479         struct ldlm_resource *res;
2480         struct ldlm_namespace *ns;
2481         ENTRY;
2482
2483         lock_res_and_lock(lock);
2484
2485         res = lock->l_resource;
2486         ns  = ldlm_res_to_ns(res);
2487
2488         /* Please do not, no matter how tempting, remove this LBUG without
2489          * talking to me first. -phik */
2490         if (lock->l_readers || lock->l_writers) {
2491                 LDLM_ERROR(lock, "lock still has references");
2492                 unlock_res_and_lock(lock);
2493                 LBUG();
2494         }
2495
2496         if (ldlm_is_waited(lock))
2497                 ldlm_del_waiting_lock(lock);
2498
2499         /* Releases cancel callback. */
2500         ldlm_cancel_callback(lock);
2501
2502         /* Yes, second time, just in case it was added again while we were
2503          * running with no res lock in ldlm_cancel_callback */
2504         if (ldlm_is_waited(lock))
2505                 ldlm_del_waiting_lock(lock);
2506
2507         ldlm_resource_unlink_lock(lock);
2508         ldlm_lock_destroy_nolock(lock);
2509
2510         if (ldlm_is_granted(lock))
2511                 ldlm_pool_del(&ns->ns_pool, lock);
2512
2513         /* Make sure we will not be called again for same lock what is possible
2514          * if not to zero out lock->l_granted_mode */
2515         lock->l_granted_mode = LCK_MINMODE;
2516         unlock_res_and_lock(lock);
2517
2518         EXIT;
2519 }
2520 EXPORT_SYMBOL(ldlm_lock_cancel);
2521
2522 /**
2523  * Set opaque data into the lock that only makes sense to upper layer.
2524  */
2525 int ldlm_lock_set_data(const struct lustre_handle *lockh, void *data)
2526 {
2527         struct ldlm_lock *lock = ldlm_handle2lock(lockh);
2528         int rc = -EINVAL;
2529         ENTRY;
2530
2531         if (lock) {
2532                 if (lock->l_ast_data == NULL)
2533                         lock->l_ast_data = data;
2534                 if (lock->l_ast_data == data)
2535                         rc = 0;
2536                 LDLM_LOCK_PUT(lock);
2537         }
2538         RETURN(rc);
2539 }
2540 EXPORT_SYMBOL(ldlm_lock_set_data);
2541
2542 struct export_cl_data {
2543         const struct lu_env     *ecl_env;
2544         struct obd_export       *ecl_exp;
2545         int                     ecl_loop;
2546 };
2547
2548 static void ldlm_cancel_lock_for_export(struct obd_export *exp,
2549                                         struct ldlm_lock *lock,
2550                                         struct export_cl_data *ecl)
2551 {
2552         struct ldlm_resource *res;
2553
2554         res = ldlm_resource_getref(lock->l_resource);
2555
2556         ldlm_lvbo_update(res, lock, NULL, 1);
2557         ldlm_lock_cancel(lock);
2558         if (!exp->exp_obd->obd_stopping)
2559                 ldlm_reprocess_all(res, lock);
2560         ldlm_resource_putref(res);
2561
2562         ecl->ecl_loop++;
2563         if ((ecl->ecl_loop & -ecl->ecl_loop) == ecl->ecl_loop) {
2564                 CDEBUG(D_INFO, "Export %p, %d locks cancelled.\n",
2565                        exp, ecl->ecl_loop);
2566         }
2567 }
2568
2569 /**
2570  * Iterator function for ldlm_export_cancel_locks.
2571  * Cancels passed locks.
2572  */
2573 static int
2574 ldlm_cancel_locks_for_export_cb(struct cfs_hash *hs, struct cfs_hash_bd *bd,
2575                                 struct hlist_node *hnode, void *data)
2576
2577 {
2578         struct export_cl_data   *ecl = (struct export_cl_data *)data;
2579         struct obd_export       *exp  = ecl->ecl_exp;
2580         struct ldlm_lock        *lock = cfs_hash_object(hs, hnode);
2581
2582         LDLM_LOCK_GET(lock);
2583         ldlm_cancel_lock_for_export(exp, lock, ecl);
2584         LDLM_LOCK_RELEASE(lock);
2585
2586         return 0;
2587 }
2588
2589 /**
2590  * Cancel all blocked locks for given export.
2591  *
2592  * Typically called on client disconnection/eviction
2593  */
2594 int ldlm_export_cancel_blocked_locks(struct obd_export *exp)
2595 {
2596         struct lu_env env;
2597         struct export_cl_data   ecl = {
2598                 .ecl_exp        = exp,
2599                 .ecl_loop       = 0,
2600         };
2601         int rc;
2602
2603         rc = lu_env_init(&env, LCT_DT_THREAD);
2604         if (rc)
2605                 RETURN(rc);
2606         ecl.ecl_env = &env;
2607
2608         while (!list_empty(&exp->exp_bl_list)) {
2609                 struct ldlm_lock *lock;
2610
2611                 spin_lock_bh(&exp->exp_bl_list_lock);
2612                 if (!list_empty(&exp->exp_bl_list)) {
2613                         lock = list_entry(exp->exp_bl_list.next,
2614                                           struct ldlm_lock, l_exp_list);
2615                         LDLM_LOCK_GET(lock);
2616                         list_del_init(&lock->l_exp_list);
2617                 } else {
2618                         lock = NULL;
2619                 }
2620                 spin_unlock_bh(&exp->exp_bl_list_lock);
2621
2622                 if (lock == NULL)
2623                         break;
2624
2625                 ldlm_cancel_lock_for_export(exp, lock, &ecl);
2626                 LDLM_LOCK_RELEASE(lock);
2627         }
2628
2629         lu_env_fini(&env);
2630
2631         CDEBUG(D_DLMTRACE, "Export %p, canceled %d locks, "
2632                "left on hash table %d.\n", exp, ecl.ecl_loop,
2633                atomic_read(&exp->exp_lock_hash->hs_count));
2634
2635         return ecl.ecl_loop;
2636 }
2637
2638 /**
2639  * Cancel all locks for given export.
2640  *
2641  * Typically called after client disconnection/eviction
2642  */
2643 int ldlm_export_cancel_locks(struct obd_export *exp)
2644 {
2645         struct export_cl_data ecl;
2646         struct lu_env env;
2647         int rc;
2648
2649         rc = lu_env_init(&env, LCT_DT_THREAD);
2650         if (rc)
2651                 RETURN(rc);
2652         ecl.ecl_env = &env;
2653         ecl.ecl_exp = exp;
2654         ecl.ecl_loop = 0;
2655
2656         cfs_hash_for_each_empty(exp->exp_lock_hash,
2657                                 ldlm_cancel_locks_for_export_cb, &ecl);
2658
2659         CDEBUG(D_DLMTRACE, "Export %p, canceled %d locks, "
2660                "left on hash table %d.\n", exp, ecl.ecl_loop,
2661                atomic_read(&exp->exp_lock_hash->hs_count));
2662
2663         if (ecl.ecl_loop > 0 &&
2664             atomic_read(&exp->exp_lock_hash->hs_count) == 0 &&
2665             exp->exp_obd->obd_stopping)
2666                 ldlm_reprocess_recovery_done(exp->exp_obd->obd_namespace);
2667
2668         lu_env_fini(&env);
2669
2670         return ecl.ecl_loop;
2671 }
2672
2673 /**
2674  * Downgrade an PW/EX lock to COS | CR mode.
2675  *
2676  * A lock mode convertion from PW/EX mode to less conflict mode. The
2677  * convertion may fail if lock was canceled before downgrade, but it doesn't
2678  * indicate any problem, because such lock has no reader or writer, and will
2679  * be released soon.
2680  *
2681  * Used by Commit on Sharing (COS) code to force object changes commit in case
2682  * of conflict. Converted lock is considered as new lock and all blocking AST
2683  * things are cleared, so any pending or new blocked lock on that lock will
2684  * cause new call to blocking_ast and force resource object commit.
2685  *
2686  * Also used by layout_change to replace EX lock to CR lock.
2687  *
2688  * \param lock A lock to convert
2689  * \param new_mode new lock mode
2690  */
2691 void ldlm_lock_mode_downgrade(struct ldlm_lock *lock, enum ldlm_mode new_mode)
2692 {
2693 #ifdef HAVE_SERVER_SUPPORT
2694         ENTRY;
2695
2696         LASSERT(new_mode == LCK_COS || new_mode == LCK_CR);
2697
2698         lock_res_and_lock(lock);
2699
2700         if (!(lock->l_granted_mode & (LCK_PW | LCK_EX))) {
2701                 unlock_res_and_lock(lock);
2702
2703                 LASSERT(lock->l_granted_mode == LCK_MINMODE);
2704                 LDLM_DEBUG(lock, "lock was canceled before downgrade");
2705                 RETURN_EXIT;
2706         }
2707
2708         ldlm_resource_unlink_lock(lock);
2709         /*
2710          * Remove the lock from pool as it will be added again in
2711          * ldlm_grant_lock() called below.
2712          */
2713         ldlm_pool_del(&ldlm_lock_to_ns(lock)->ns_pool, lock);
2714
2715         /* Consider downgraded lock as a new lock and clear all states
2716          * related to a previous blocking AST processing.
2717          */
2718         ldlm_clear_blocking_data(lock);
2719
2720         lock->l_req_mode = new_mode;
2721         ldlm_grant_lock(lock, NULL);
2722         unlock_res_and_lock(lock);
2723
2724         ldlm_reprocess_all(lock->l_resource, lock);
2725
2726         EXIT;
2727 #endif
2728 }
2729 EXPORT_SYMBOL(ldlm_lock_mode_downgrade);
2730
2731 /**
2732  * Print lock with lock handle \a lockh description into debug log.
2733  *
2734  * Used when printing all locks on a resource for debug purposes.
2735  */
2736 void ldlm_lock_dump_handle(int level, const struct lustre_handle *lockh)
2737 {
2738         struct ldlm_lock *lock;
2739
2740         if (!((libcfs_debug | D_ERROR) & level))
2741                 return;
2742
2743         lock = ldlm_handle2lock(lockh);
2744         if (lock == NULL)
2745                 return;
2746
2747         LDLM_DEBUG_LIMIT(level, lock, "###");
2748
2749         LDLM_LOCK_PUT(lock);
2750 }
2751 EXPORT_SYMBOL(ldlm_lock_dump_handle);
2752
2753 /**
2754  * Print lock information with custom message into debug log.
2755  * Helper function.
2756  */
2757 void _ldlm_lock_debug(struct ldlm_lock *lock,
2758                       struct libcfs_debug_msg_data *msgdata,
2759                       const char *fmt, ...)
2760 {
2761         va_list args;
2762         struct obd_export *exp = lock->l_export;
2763         struct ldlm_resource *resource = NULL;
2764         struct va_format vaf;
2765         char *nid = "local";
2766
2767         rcu_read_lock();
2768         resource = rcu_dereference(lock->l_resource);
2769         if (resource && !atomic_inc_not_zero(&resource->lr_refcount))
2770                 resource = NULL;
2771         rcu_read_unlock();
2772
2773         va_start(args, fmt);
2774         vaf.fmt = fmt;
2775         vaf.va = &args;
2776
2777         if (exp && exp->exp_connection) {
2778                 nid = obd_export_nid2str(exp);
2779         } else if (exp && exp->exp_obd != NULL) {
2780                 struct obd_import *imp = exp->exp_obd->u.cli.cl_import;
2781                 nid = obd_import_nid2str(imp);
2782         }
2783
2784         if (resource == NULL) {
2785                 libcfs_debug_msg(msgdata,
2786                                  "%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",
2787                                  &vaf,
2788                                  lock,
2789                                  lock->l_handle.h_cookie,
2790                                  refcount_read(&lock->l_handle.h_ref),
2791                                  lock->l_readers, lock->l_writers,
2792                                  ldlm_lockname[lock->l_granted_mode],
2793                                  ldlm_lockname[lock->l_req_mode],
2794                                  lock->l_flags, nid,
2795                                  lock->l_remote_handle.cookie,
2796                                  exp ? refcount_read(&exp->exp_handle.h_ref) : -99,
2797                                  lock->l_pid, lock->l_callback_timestamp,
2798                                  lock->l_lvb_type);
2799                 va_end(args);
2800                 return;
2801         }
2802
2803         switch (resource->lr_type) {
2804         case LDLM_EXTENT:
2805                 libcfs_debug_msg(msgdata,
2806                                  "%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",
2807                                  &vaf,
2808                                  ldlm_lock_to_ns_name(lock), lock,
2809                                  lock->l_handle.h_cookie,
2810                                  refcount_read(&lock->l_handle.h_ref),
2811                                  lock->l_readers, lock->l_writers,
2812                                  ldlm_lockname[lock->l_granted_mode],
2813                                  ldlm_lockname[lock->l_req_mode],
2814                                  PLDLMRES(resource),
2815                                  atomic_read(&resource->lr_refcount),
2816                                  ldlm_typename[resource->lr_type],
2817                                  lock->l_policy_data.l_extent.start,
2818                                  lock->l_policy_data.l_extent.end,
2819                                  lock->l_req_extent.start, lock->l_req_extent.end,
2820                                  lock->l_req_extent.gid,
2821                                  lock->l_flags, nid,
2822                                  lock->l_remote_handle.cookie,
2823                                  exp ? refcount_read(&exp->exp_handle.h_ref) : -99,
2824                                  lock->l_pid, lock->l_callback_timestamp,
2825                                  lock->l_lvb_type);
2826                 break;
2827
2828         case LDLM_FLOCK:
2829                 libcfs_debug_msg(msgdata,
2830                                  "%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",
2831                                  &vaf,
2832                                  ldlm_lock_to_ns_name(lock), lock,
2833                                  lock->l_handle.h_cookie,
2834                                  refcount_read(&lock->l_handle.h_ref),
2835                                  lock->l_readers, lock->l_writers,
2836                                  ldlm_lockname[lock->l_granted_mode],
2837                                  ldlm_lockname[lock->l_req_mode],
2838                                  PLDLMRES(resource),
2839                                  atomic_read(&resource->lr_refcount),
2840                                  ldlm_typename[resource->lr_type],
2841                                  lock->l_policy_data.l_flock.pid,
2842                                  lock->l_policy_data.l_flock.start,
2843                                  lock->l_policy_data.l_flock.end,
2844                                  lock->l_flags, nid,
2845                                  lock->l_remote_handle.cookie,
2846                                  exp ? refcount_read(&exp->exp_handle.h_ref) : -99,
2847                                  lock->l_pid, lock->l_callback_timestamp);
2848                 break;
2849
2850         case LDLM_IBITS:
2851                 libcfs_debug_msg(msgdata,
2852                                  "%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",
2853                                  &vaf,
2854                                  ldlm_lock_to_ns_name(lock),
2855                                  lock, lock->l_handle.h_cookie,
2856                                  refcount_read(&lock->l_handle.h_ref),
2857                                  lock->l_readers, lock->l_writers,
2858                                  ldlm_lockname[lock->l_granted_mode],
2859                                  ldlm_lockname[lock->l_req_mode],
2860                                  PLDLMRES(resource),
2861                                  lock->l_policy_data.l_inodebits.bits,
2862                                  lock->l_policy_data.l_inodebits.try_bits,
2863                                  atomic_read(&resource->lr_refcount),
2864                                  ldlm_typename[resource->lr_type],
2865                                  lock->l_policy_data.l_inodebits.li_gid,
2866                                  lock->l_flags, nid,
2867                                  lock->l_remote_handle.cookie,
2868                                  exp ? refcount_read(&exp->exp_handle.h_ref) : -99,
2869                                  lock->l_pid, lock->l_callback_timestamp,
2870                                  lock->l_lvb_type);
2871                 break;
2872
2873         default:
2874                 libcfs_debug_msg(msgdata,
2875                                  "%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",
2876                                  &vaf,
2877                                  ldlm_lock_to_ns_name(lock),
2878                                  lock, lock->l_handle.h_cookie,
2879                                  refcount_read(&lock->l_handle.h_ref),
2880                                  lock->l_readers, lock->l_writers,
2881                                  ldlm_lockname[lock->l_granted_mode],
2882                                  ldlm_lockname[lock->l_req_mode],
2883                                  PLDLMRES(resource),
2884                                  atomic_read(&resource->lr_refcount),
2885                                  ldlm_typename[resource->lr_type],
2886                                  lock->l_flags, nid,
2887                                  lock->l_remote_handle.cookie,
2888                                  exp ? refcount_read(&exp->exp_handle.h_ref) : -99,
2889                                  lock->l_pid, lock->l_callback_timestamp,
2890                                  lock->l_lvb_type);
2891                 break;
2892         }
2893         va_end(args);
2894         ldlm_resource_putref(resource);
2895 }
2896 EXPORT_SYMBOL(_ldlm_lock_debug);