Whamcloud - gitweb
LU-13264 osc: ensure lu_ref work atomic from osc_lock_upcall()
[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         spin_lock_init(&lock->l_lock);
473         lock->l_resource = resource;
474         lu_ref_add(&resource->lr_reference, "lock", lock);
475
476         refcount_set(&lock->l_handle.h_ref, 2);
477         INIT_LIST_HEAD(&lock->l_res_link);
478         INIT_LIST_HEAD(&lock->l_lru);
479         INIT_LIST_HEAD(&lock->l_pending_chain);
480         INIT_LIST_HEAD(&lock->l_bl_ast);
481         INIT_LIST_HEAD(&lock->l_cp_ast);
482         INIT_LIST_HEAD(&lock->l_rk_ast);
483         init_waitqueue_head(&lock->l_waitq);
484         lock->l_blocking_lock = NULL;
485         INIT_LIST_HEAD(&lock->l_sl_mode);
486         INIT_LIST_HEAD(&lock->l_sl_policy);
487         INIT_HLIST_NODE(&lock->l_exp_hash);
488         INIT_HLIST_NODE(&lock->l_exp_flock_hash);
489
490         lprocfs_counter_incr(ldlm_res_to_ns(resource)->ns_stats,
491                              LDLM_NSS_LOCKS);
492         INIT_HLIST_NODE(&lock->l_handle.h_link);
493         class_handle_hash(&lock->l_handle, lock_handle_owner);
494
495         lu_ref_init(&lock->l_reference);
496         lu_ref_add(&lock->l_reference, "hash", lock);
497         lock->l_callback_timeout = 0;
498         lock->l_activity = 0;
499
500 #if LUSTRE_TRACKS_LOCK_EXP_REFS
501         INIT_LIST_HEAD(&lock->l_exp_refs_link);
502         lock->l_exp_refs_nr = 0;
503         lock->l_exp_refs_target = NULL;
504 #endif
505         INIT_LIST_HEAD(&lock->l_exp_list);
506
507         RETURN(lock);
508 }
509
510 /**
511  * Moves LDLM lock \a lock to another resource.
512  * This is used on client when server returns some other lock than requested
513  * (typically as a result of intent operation)
514  */
515 int ldlm_lock_change_resource(struct ldlm_namespace *ns, struct ldlm_lock *lock,
516                               const struct ldlm_res_id *new_resid)
517 {
518         struct ldlm_resource *oldres = lock->l_resource;
519         struct ldlm_resource *newres;
520         int type;
521         ENTRY;
522
523         LASSERT(ns_is_client(ns));
524
525         lock_res_and_lock(lock);
526         if (memcmp(new_resid, &lock->l_resource->lr_name,
527                    sizeof(lock->l_resource->lr_name)) == 0) {
528                 /* Nothing to do */
529                 unlock_res_and_lock(lock);
530                 RETURN(0);
531         }
532
533         LASSERT(new_resid->name[0] != 0);
534
535         /* This function assumes that the lock isn't on any lists */
536         LASSERT(list_empty(&lock->l_res_link));
537
538         type = oldres->lr_type;
539         unlock_res_and_lock(lock);
540
541         newres = ldlm_resource_get(ns, NULL, new_resid, type, 1);
542         if (IS_ERR(newres))
543                 RETURN(PTR_ERR(newres));
544
545         lu_ref_add(&newres->lr_reference, "lock", lock);
546         /*
547          * To flip the lock from the old to the new resource, lock, oldres and
548          * newres have to be locked. Resource spin-locks are nested within
549          * lock->l_lock, and are taken in the memory address order to avoid
550          * dead-locks.
551          */
552         spin_lock(&lock->l_lock);
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         lock->l_resource = newres;
564         unlock_res(oldres);
565         unlock_res_and_lock(lock);
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", current);
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", current);
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         ENTRY;
849
850         lock_res_and_lock(lock);
851
852         ns = ldlm_lock_to_ns(lock);
853
854         ldlm_lock_decref_internal_nolock(lock, mode);
855
856         if ((ldlm_is_local(lock) || lock->l_req_mode == LCK_GROUP) &&
857             !lock->l_readers && !lock->l_writers) {
858                 /* If this is a local lock on a server namespace and this was
859                  * the last reference, cancel the lock.
860                  *
861                  * Group locks are special:
862                  * They must not go in LRU, but they are not called back
863                  * like non-group locks, instead they are manually released.
864                  * They have an l_writers reference which they keep until
865                  * they are manually released, so we remove them when they have
866                  * no more reader or writer references. - LU-6368 */
867                 ldlm_set_cbpending(lock);
868         }
869
870         if (!lock->l_readers && !lock->l_writers && ldlm_is_cbpending(lock)) {
871                 /* If we received a blocked AST and this was the last reference,
872                  * run the callback. */
873                 if (ldlm_is_ns_srv(lock) && lock->l_export)
874                         CERROR("FL_CBPENDING set on non-local lock--just a "
875                                "warning\n");
876
877                 LDLM_DEBUG(lock, "final decref done on cbpending lock");
878
879                 LDLM_LOCK_GET(lock); /* dropped by bl thread */
880                 ldlm_lock_remove_from_lru(lock);
881                 unlock_res_and_lock(lock);
882
883                 if (ldlm_is_fail_loc(lock))
884                         OBD_RACE(OBD_FAIL_LDLM_CP_BL_RACE);
885
886                 if (ldlm_is_atomic_cb(lock) ||
887                     ldlm_bl_to_thread_lock(ns, NULL, lock) != 0)
888                         ldlm_handle_bl_callback(ns, NULL, lock);
889         } else if (ns_is_client(ns) &&
890                    !lock->l_readers && !lock->l_writers &&
891                    !ldlm_is_no_lru(lock) &&
892                    !ldlm_is_bl_ast(lock) &&
893                    !ldlm_is_converting(lock)) {
894
895                 LDLM_DEBUG(lock, "add lock into lru list");
896
897                 /* If this is a client-side namespace and this was the last
898                  * reference, put it on the LRU. */
899                 ldlm_lock_add_to_lru(lock);
900                 unlock_res_and_lock(lock);
901
902                 if (ldlm_is_fail_loc(lock))
903                         OBD_RACE(OBD_FAIL_LDLM_CP_BL_RACE);
904
905                 /* Call ldlm_cancel_lru() only if EARLY_CANCEL and LRU RESIZE
906                  * are not supported by the server, otherwise, it is done on
907                  * enqueue. */
908                 if (!exp_connect_cancelset(lock->l_conn_export) &&
909                     !ns_connect_lru_resize(ns))
910                         ldlm_cancel_lru(ns, 0, LCF_ASYNC, 0);
911         } else {
912                 LDLM_DEBUG(lock, "do not add lock into lru list");
913                 unlock_res_and_lock(lock);
914         }
915
916         EXIT;
917 }
918
919 /**
920  * Decrease reader/writer refcount for LDLM lock with handle \a lockh
921  */
922 void ldlm_lock_decref(const struct lustre_handle *lockh, enum ldlm_mode mode)
923 {
924         struct ldlm_lock *lock = __ldlm_handle2lock(lockh, 0);
925         LASSERTF(lock != NULL, "Non-existing lock: %#llx\n", lockh->cookie);
926         ldlm_lock_decref_internal(lock, mode);
927         LDLM_LOCK_PUT(lock);
928 }
929 EXPORT_SYMBOL(ldlm_lock_decref);
930
931 /**
932  * Decrease reader/writer refcount for LDLM lock with handle
933  * \a lockh and mark it for subsequent cancellation once r/w refcount
934  * drops to zero instead of putting into LRU.
935  *
936  */
937 void ldlm_lock_decref_and_cancel(const struct lustre_handle *lockh,
938                                  enum ldlm_mode mode)
939 {
940         struct ldlm_lock *lock = __ldlm_handle2lock(lockh, 0);
941         ENTRY;
942
943         LASSERT(lock != NULL);
944
945         LDLM_DEBUG(lock, "ldlm_lock_decref(%s)", ldlm_lockname[mode]);
946         lock_res_and_lock(lock);
947         ldlm_set_cbpending(lock);
948         unlock_res_and_lock(lock);
949         ldlm_lock_decref_internal(lock, mode);
950         LDLM_LOCK_PUT(lock);
951 }
952 EXPORT_SYMBOL(ldlm_lock_decref_and_cancel);
953
954 struct sl_insert_point {
955         struct list_head *res_link;
956         struct list_head *mode_link;
957         struct list_head *policy_link;
958 };
959
960 /**
961  * Finds a position to insert the new lock into granted lock list.
962  *
963  * Used for locks eligible for skiplist optimization.
964  *
965  * Parameters:
966  *      queue [input]:  the granted list where search acts on;
967  *      req [input]:    the lock whose position to be located;
968  *      prev [output]:  positions within 3 lists to insert @req to
969  * Return Value:
970  *      filled @prev
971  * NOTE: called by
972  *  - ldlm_grant_lock_with_skiplist
973  */
974 static void search_granted_lock(struct list_head *queue,
975                                 struct ldlm_lock *req,
976                                 struct sl_insert_point *prev)
977 {
978         struct list_head *tmp;
979         struct ldlm_lock *lock, *mode_end, *policy_end;
980         ENTRY;
981
982         list_for_each(tmp, queue) {
983                 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
984
985                 mode_end = list_entry(lock->l_sl_mode.prev,
986                                           struct ldlm_lock, l_sl_mode);
987
988                 if (lock->l_req_mode != req->l_req_mode) {
989                         /* jump to last lock of mode group */
990                         tmp = &mode_end->l_res_link;
991                         continue;
992                 }
993
994                 /* suitable mode group is found */
995                 if (lock->l_resource->lr_type == LDLM_PLAIN) {
996                         /* insert point is last lock of the mode group */
997                         prev->res_link = &mode_end->l_res_link;
998                         prev->mode_link = &mode_end->l_sl_mode;
999                         prev->policy_link = &req->l_sl_policy;
1000                         EXIT;
1001                         return;
1002                 } else if (lock->l_resource->lr_type == LDLM_IBITS) {
1003                         for (;;) {
1004                                 policy_end =
1005                                         list_entry(lock->l_sl_policy.prev,
1006                                                        struct ldlm_lock,
1007                                                        l_sl_policy);
1008
1009                                 if (lock->l_policy_data.l_inodebits.bits ==
1010                                     req->l_policy_data.l_inodebits.bits) {
1011                                         /* insert point is last lock of
1012                                          * the policy group */
1013                                         prev->res_link =
1014                                                 &policy_end->l_res_link;
1015                                         prev->mode_link =
1016                                                 &policy_end->l_sl_mode;
1017                                         prev->policy_link =
1018                                                 &policy_end->l_sl_policy;
1019                                         EXIT;
1020                                         return;
1021                                 }
1022
1023                                 if (policy_end == mode_end)
1024                                         /* done with mode group */
1025                                         break;
1026
1027                                 /* go to next policy group within mode group */
1028                                 tmp = policy_end->l_res_link.next;
1029                                 lock = list_entry(tmp, struct ldlm_lock,
1030                                                       l_res_link);
1031                         }  /* loop over policy groups within the mode group */
1032
1033                         /* insert point is last lock of the mode group,
1034                          * new policy group is started */
1035                         prev->res_link = &mode_end->l_res_link;
1036                         prev->mode_link = &mode_end->l_sl_mode;
1037                         prev->policy_link = &req->l_sl_policy;
1038                         EXIT;
1039                         return;
1040                 } else {
1041                         LDLM_ERROR(lock,"is not LDLM_PLAIN or LDLM_IBITS lock");
1042                         LBUG();
1043                 }
1044         }
1045
1046         /* insert point is last lock on the queue,
1047          * new mode group and new policy group are started */
1048         prev->res_link = queue->prev;
1049         prev->mode_link = &req->l_sl_mode;
1050         prev->policy_link = &req->l_sl_policy;
1051         EXIT;
1052 }
1053
1054 /**
1055  * Add a lock into resource granted list after a position described by
1056  * \a prev.
1057  */
1058 static void ldlm_granted_list_add_lock(struct ldlm_lock *lock,
1059                                        struct sl_insert_point *prev)
1060 {
1061         struct ldlm_resource *res = lock->l_resource;
1062         ENTRY;
1063
1064         check_res_locked(res);
1065
1066         ldlm_resource_dump(D_INFO, res);
1067         LDLM_DEBUG(lock, "About to add lock:");
1068
1069         if (ldlm_is_destroyed(lock)) {
1070                 CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n");
1071                 return;
1072         }
1073
1074         LASSERT(list_empty(&lock->l_res_link));
1075         LASSERT(list_empty(&lock->l_sl_mode));
1076         LASSERT(list_empty(&lock->l_sl_policy));
1077
1078         /*
1079          * lock->link == prev->link means lock is first starting the group.
1080          * Don't re-add to itself to suppress kernel warnings.
1081          */
1082         if (&lock->l_res_link != prev->res_link)
1083                 list_add(&lock->l_res_link, prev->res_link);
1084         if (&lock->l_sl_mode != prev->mode_link)
1085                 list_add(&lock->l_sl_mode, prev->mode_link);
1086         if (&lock->l_sl_policy != prev->policy_link)
1087                 list_add(&lock->l_sl_policy, prev->policy_link);
1088
1089         EXIT;
1090 }
1091
1092 /**
1093  * Add a lock to granted list on a resource maintaining skiplist
1094  * correctness.
1095  */
1096 void ldlm_grant_lock_with_skiplist(struct ldlm_lock *lock)
1097 {
1098         struct sl_insert_point prev;
1099
1100         LASSERT(ldlm_is_granted(lock));
1101
1102         search_granted_lock(&lock->l_resource->lr_granted, lock, &prev);
1103         ldlm_granted_list_add_lock(lock, &prev);
1104 }
1105
1106 /**
1107  * Perform lock granting bookkeeping.
1108  *
1109  * Includes putting the lock into granted list and updating lock mode.
1110  * NOTE: called by
1111  *  - ldlm_lock_enqueue
1112  *  - ldlm_reprocess_queue
1113  *
1114  * must be called with lr_lock held
1115  */
1116 void ldlm_grant_lock(struct ldlm_lock *lock, struct list_head *work_list)
1117 {
1118         struct ldlm_resource *res = lock->l_resource;
1119         ENTRY;
1120
1121         check_res_locked(res);
1122
1123         lock->l_granted_mode = lock->l_req_mode;
1124
1125         if (work_list && lock->l_completion_ast != NULL)
1126                 ldlm_add_ast_work_item(lock, NULL, work_list);
1127
1128         if (res->lr_type == LDLM_PLAIN || res->lr_type == LDLM_IBITS)
1129                 ldlm_grant_lock_with_skiplist(lock);
1130         else if (res->lr_type == LDLM_EXTENT)
1131                 ldlm_extent_add_lock(res, lock);
1132         else if (res->lr_type == LDLM_FLOCK) {
1133                 /* We should not add locks to granted list in the following
1134                  * cases:
1135                  * - this is an UNLOCK but not a real lock;
1136                  * - this is a TEST lock;
1137                  * - this is a F_CANCELLK lock (async flock has req_mode == 0)
1138                  * - this is a deadlock (flock cannot be granted) */
1139                 if (lock->l_req_mode == 0 ||
1140                     lock->l_req_mode == LCK_NL ||
1141                     ldlm_is_test_lock(lock) ||
1142                     ldlm_is_flock_deadlock(lock))
1143                         RETURN_EXIT;
1144                 ldlm_resource_add_lock(res, &res->lr_granted, lock);
1145         } else {
1146                 LBUG();
1147         }
1148
1149         ldlm_pool_add(&ldlm_res_to_ns(res)->ns_pool, lock);
1150         EXIT;
1151 }
1152
1153 /**
1154  * Check if the given @lock meets the criteria for a match.
1155  * A reference on the lock is taken if matched.
1156  *
1157  * \param lock     test-against this lock
1158  * \param data     parameters
1159  */
1160 static int 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 INTERVAL_ITER_STOP;
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 INTERVAL_ITER_CONT;
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 INTERVAL_ITER_CONT;
1182         if (!data->lmd_unref && ldlm_is_cbpending(lock) &&
1183             lock->l_readers == 0 && lock->l_writers == 0)
1184                 return INTERVAL_ITER_CONT;
1185
1186         if (!(lock->l_req_mode & *data->lmd_mode))
1187                 return INTERVAL_ITER_CONT;
1188
1189         /* When we search for ast_data, we are not doing a traditional match,
1190          * so we don't worry about IBITS or extent matching.
1191          */
1192         if (data->lmd_has_ast_data) {
1193                 if (!lock->l_ast_data)
1194                         return INTERVAL_ITER_CONT;
1195
1196                 goto matched;
1197         }
1198
1199         match = lock->l_req_mode;
1200
1201         switch (lock->l_resource->lr_type) {
1202         case LDLM_EXTENT:
1203                 if (lpol->l_extent.start > data->lmd_policy->l_extent.start ||
1204                     lpol->l_extent.end < data->lmd_policy->l_extent.end)
1205                         return INTERVAL_ITER_CONT;
1206
1207                 if (unlikely(match == LCK_GROUP) &&
1208                     data->lmd_policy->l_extent.gid != LDLM_GID_ANY &&
1209                     lpol->l_extent.gid != data->lmd_policy->l_extent.gid)
1210                         return INTERVAL_ITER_CONT;
1211                 break;
1212         case LDLM_IBITS:
1213                 /* We match if we have existing lock with same or wider set
1214                    of bits. */
1215                 if ((lpol->l_inodebits.bits &
1216                      data->lmd_policy->l_inodebits.bits) !=
1217                     data->lmd_policy->l_inodebits.bits)
1218                         return INTERVAL_ITER_CONT;
1219                 break;
1220         default:
1221                 ;
1222         }
1223
1224         /* We match if we have existing lock with same or wider set
1225            of bits. */
1226         if (!data->lmd_unref && LDLM_HAVE_MASK(lock, GONE))
1227                 return INTERVAL_ITER_CONT;
1228
1229         if (!equi(data->lmd_flags & LDLM_FL_LOCAL_ONLY, ldlm_is_local(lock)))
1230                 return INTERVAL_ITER_CONT;
1231
1232         /* Filter locks by skipping flags */
1233         if (data->lmd_skip_flags & lock->l_flags)
1234                 return INTERVAL_ITER_CONT;
1235
1236 matched:
1237         if (data->lmd_flags & LDLM_FL_TEST_LOCK) {
1238                 LDLM_LOCK_GET(lock);
1239                 ldlm_lock_touch_in_lru(lock);
1240         } else {
1241                 ldlm_lock_addref_internal_nolock(lock, match);
1242         }
1243
1244         *data->lmd_mode = match;
1245         data->lmd_lock = lock;
1246
1247         return INTERVAL_ITER_STOP;
1248 }
1249
1250 static unsigned int itree_overlap_cb(struct interval_node *in, void *args)
1251 {
1252         struct ldlm_interval *node = to_ldlm_interval(in);
1253         struct ldlm_match_data *data = args;
1254         struct ldlm_lock *lock;
1255         int rc;
1256
1257         list_for_each_entry(lock, &node->li_group, l_sl_policy) {
1258                 rc = lock_matches(lock, data);
1259                 if (rc == INTERVAL_ITER_STOP)
1260                         return INTERVAL_ITER_STOP;
1261         }
1262         return INTERVAL_ITER_CONT;
1263 }
1264
1265 /**
1266  * Search for a lock with given parameters in interval trees.
1267  *
1268  * \param res      search for a lock in this resource
1269  * \param data     parameters
1270  *
1271  * \retval a referenced lock or NULL.
1272  */
1273 struct ldlm_lock *search_itree(struct ldlm_resource *res,
1274                                struct ldlm_match_data *data)
1275 {
1276         struct interval_node_extent ext = {
1277                 .start     = data->lmd_policy->l_extent.start,
1278                 .end       = data->lmd_policy->l_extent.end
1279         };
1280         int idx;
1281
1282         data->lmd_lock = NULL;
1283
1284         for (idx = 0; idx < LCK_MODE_NUM; idx++) {
1285                 struct ldlm_interval_tree *tree = &res->lr_itree[idx];
1286
1287                 if (tree->lit_root == NULL)
1288                         continue;
1289
1290                 if (!(tree->lit_mode & *data->lmd_mode))
1291                         continue;
1292
1293                 interval_search(tree->lit_root, &ext,
1294                                 itree_overlap_cb, data);
1295                 if (data->lmd_lock)
1296                         return data->lmd_lock;
1297         }
1298
1299         return NULL;
1300 }
1301 EXPORT_SYMBOL(search_itree);
1302
1303
1304 /**
1305  * Search for a lock with given properties in a queue.
1306  *
1307  * \param queue    search for a lock in this queue
1308  * \param data     parameters
1309  *
1310  * \retval a referenced lock or NULL.
1311  */
1312 static struct ldlm_lock *search_queue(struct list_head *queue,
1313                                       struct ldlm_match_data *data)
1314 {
1315         struct ldlm_lock *lock;
1316         int rc;
1317
1318         data->lmd_lock = NULL;
1319
1320         list_for_each_entry(lock, queue, l_res_link) {
1321                 rc = lock_matches(lock, data);
1322                 if (rc == INTERVAL_ITER_STOP)
1323                         return data->lmd_lock;
1324         }
1325
1326         return NULL;
1327 }
1328
1329 void ldlm_lock_fail_match_locked(struct ldlm_lock *lock)
1330 {
1331         if ((lock->l_flags & LDLM_FL_FAIL_NOTIFIED) == 0) {
1332                 lock->l_flags |= LDLM_FL_FAIL_NOTIFIED;
1333                 wake_up_all(&lock->l_waitq);
1334         }
1335 }
1336 EXPORT_SYMBOL(ldlm_lock_fail_match_locked);
1337
1338 void ldlm_lock_fail_match(struct ldlm_lock *lock)
1339 {
1340         lock_res_and_lock(lock);
1341         ldlm_lock_fail_match_locked(lock);
1342         unlock_res_and_lock(lock);
1343 }
1344
1345 /**
1346  * Mark lock as "matchable" by OST.
1347  *
1348  * Used to prevent certain races in LOV/OSC where the lock is granted, but LVB
1349  * is not yet valid.
1350  * Assumes LDLM lock is already locked.
1351  */
1352 void ldlm_lock_allow_match_locked(struct ldlm_lock *lock)
1353 {
1354         ldlm_set_lvb_ready(lock);
1355         wake_up_all(&lock->l_waitq);
1356 }
1357 EXPORT_SYMBOL(ldlm_lock_allow_match_locked);
1358
1359 /**
1360  * Mark lock as "matchable" by OST.
1361  * Locks the lock and then \see ldlm_lock_allow_match_locked
1362  */
1363 void ldlm_lock_allow_match(struct ldlm_lock *lock)
1364 {
1365         lock_res_and_lock(lock);
1366         ldlm_lock_allow_match_locked(lock);
1367         unlock_res_and_lock(lock);
1368 }
1369 EXPORT_SYMBOL(ldlm_lock_allow_match);
1370
1371 /**
1372  * Attempt to find a lock with specified properties.
1373  *
1374  * Typically returns a reference to matched lock unless LDLM_FL_TEST_LOCK is
1375  * set in \a flags
1376  *
1377  * Can be called in two ways:
1378  *
1379  * If 'ns' is NULL, then lockh describes an existing lock that we want to look
1380  * for a duplicate of.
1381  *
1382  * Otherwise, all of the fields must be filled in, to match against.
1383  *
1384  * If 'flags' contains LDLM_FL_LOCAL_ONLY, then only match local locks on the
1385  *     server (ie, connh is NULL)
1386  * If 'flags' contains LDLM_FL_BLOCK_GRANTED, then only locks on the granted
1387  *     list will be considered
1388  * If 'flags' contains LDLM_FL_CBPENDING, then locks that have been marked
1389  *     to be canceled can still be matched as long as they still have reader
1390  *     or writer refernces
1391  * If 'flags' contains LDLM_FL_TEST_LOCK, then don't actually reference a lock,
1392  *     just tell us if we would have matched.
1393  *
1394  * \retval 1 if it finds an already-existing lock that is compatible; in this
1395  * case, lockh is filled in with a addref()ed lock
1396  *
1397  * We also check security context, and if that fails we simply return 0 (to
1398  * keep caller code unchanged), the context failure will be discovered by
1399  * caller sometime later.
1400  */
1401 enum ldlm_mode ldlm_lock_match_with_skip(struct ldlm_namespace *ns,
1402                                          __u64 flags, __u64 skip_flags,
1403                                          const struct ldlm_res_id *res_id,
1404                                          enum ldlm_type type,
1405                                          union ldlm_policy_data *policy,
1406                                          enum ldlm_mode mode,
1407                                          struct lustre_handle *lockh, int unref)
1408 {
1409         struct ldlm_match_data data = {
1410                 .lmd_old = NULL,
1411                 .lmd_lock = NULL,
1412                 .lmd_mode = &mode,
1413                 .lmd_policy = policy,
1414                 .lmd_flags = flags,
1415                 .lmd_skip_flags = skip_flags,
1416                 .lmd_unref = unref,
1417                 .lmd_has_ast_data = false,
1418         };
1419         struct ldlm_resource *res;
1420         struct ldlm_lock *lock;
1421         int matched;
1422
1423         ENTRY;
1424
1425         if (ns == NULL) {
1426                 data.lmd_old = ldlm_handle2lock(lockh);
1427                 LASSERT(data.lmd_old != NULL);
1428
1429                 ns = ldlm_lock_to_ns(data.lmd_old);
1430                 res_id = &data.lmd_old->l_resource->lr_name;
1431                 type = data.lmd_old->l_resource->lr_type;
1432                 *data.lmd_mode = data.lmd_old->l_req_mode;
1433         }
1434
1435         res = ldlm_resource_get(ns, NULL, res_id, type, 0);
1436         if (IS_ERR(res)) {
1437                 LASSERT(data.lmd_old == NULL);
1438                 RETURN(0);
1439         }
1440
1441         LDLM_RESOURCE_ADDREF(res);
1442         lock_res(res);
1443         if (res->lr_type == LDLM_EXTENT)
1444                 lock = search_itree(res, &data);
1445         else
1446                 lock = search_queue(&res->lr_granted, &data);
1447         if (!lock && !(flags & LDLM_FL_BLOCK_GRANTED))
1448                 lock = search_queue(&res->lr_waiting, &data);
1449         matched = lock ? mode : 0;
1450         unlock_res(res);
1451         LDLM_RESOURCE_DELREF(res);
1452         ldlm_resource_putref(res);
1453
1454         if (lock) {
1455                 ldlm_lock2handle(lock, lockh);
1456                 if ((flags & LDLM_FL_LVB_READY) &&
1457                     (!ldlm_is_lvb_ready(lock))) {
1458                         __u64 wait_flags = LDLM_FL_LVB_READY |
1459                                 LDLM_FL_DESTROYED | LDLM_FL_FAIL_NOTIFIED;
1460
1461                         if (lock->l_completion_ast) {
1462                                 int err = lock->l_completion_ast(lock,
1463                                                         LDLM_FL_WAIT_NOREPROC,
1464                                                         NULL);
1465                                 if (err)
1466                                         GOTO(out_fail_match, matched = 0);
1467                         }
1468
1469                         wait_event_idle_timeout(
1470                                 lock->l_waitq,
1471                                 lock->l_flags & wait_flags,
1472                                 cfs_time_seconds(obd_timeout));
1473
1474                         if (!ldlm_is_lvb_ready(lock))
1475                                 GOTO(out_fail_match, matched = 0);
1476                 }
1477
1478                 /* check user's security context */
1479                 if (lock->l_conn_export &&
1480                     sptlrpc_import_check_ctx(
1481                                 class_exp2cliimp(lock->l_conn_export)))
1482                         GOTO(out_fail_match, matched = 0);
1483
1484                 LDLM_DEBUG(lock, "matched (%llu %llu)",
1485                            (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1486                            res_id->name[2] : policy->l_extent.start,
1487                            (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1488                            res_id->name[3] : policy->l_extent.end);
1489
1490 out_fail_match:
1491                 if (flags & LDLM_FL_TEST_LOCK)
1492                         LDLM_LOCK_RELEASE(lock);
1493                 else if (!matched)
1494                         ldlm_lock_decref_internal(lock, mode);
1495         }
1496
1497         /* less verbose for test-only */
1498         if (!matched && !(flags & LDLM_FL_TEST_LOCK)) {
1499                 LDLM_DEBUG_NOLOCK("not matched ns %p type %u mode %u res "
1500                                   "%llu/%llu (%llu %llu)", ns,
1501                                   type, mode, res_id->name[0], res_id->name[1],
1502                                   (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1503                                   res_id->name[2] : policy->l_extent.start,
1504                                   (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1505                                   res_id->name[3] : policy->l_extent.end);
1506         }
1507         if (data.lmd_old != NULL)
1508                 LDLM_LOCK_PUT(data.lmd_old);
1509
1510         return matched;
1511 }
1512 EXPORT_SYMBOL(ldlm_lock_match_with_skip);
1513
1514 enum ldlm_mode ldlm_revalidate_lock_handle(const struct lustre_handle *lockh,
1515                                            __u64 *bits)
1516 {
1517         struct ldlm_lock *lock;
1518         enum ldlm_mode mode = 0;
1519         ENTRY;
1520
1521         lock = ldlm_handle2lock(lockh);
1522         if (lock != NULL) {
1523                 lock_res_and_lock(lock);
1524                 if (LDLM_HAVE_MASK(lock, GONE))
1525                         GOTO(out, mode);
1526
1527                 if (ldlm_is_cbpending(lock) &&
1528                     lock->l_readers == 0 && lock->l_writers == 0)
1529                         GOTO(out, mode);
1530
1531                 if (bits)
1532                         *bits = lock->l_policy_data.l_inodebits.bits;
1533                 mode = lock->l_granted_mode;
1534                 ldlm_lock_addref_internal_nolock(lock, mode);
1535         }
1536
1537         EXIT;
1538
1539 out:
1540         if (lock != NULL) {
1541                 unlock_res_and_lock(lock);
1542                 LDLM_LOCK_PUT(lock);
1543         }
1544         return mode;
1545 }
1546 EXPORT_SYMBOL(ldlm_revalidate_lock_handle);
1547
1548 /** The caller must guarantee that the buffer is large enough. */
1549 int ldlm_fill_lvb(struct ldlm_lock *lock, struct req_capsule *pill,
1550                   enum req_location loc, void *data, int size)
1551 {
1552         void *lvb;
1553         ENTRY;
1554
1555         LASSERT(data != NULL);
1556         LASSERT(size >= 0);
1557
1558         switch (lock->l_lvb_type) {
1559         case LVB_T_OST:
1560                 if (size == sizeof(struct ost_lvb)) {
1561                         if (loc == RCL_CLIENT)
1562                                 lvb = req_capsule_client_swab_get(pill,
1563                                                 &RMF_DLM_LVB,
1564                                                 lustre_swab_ost_lvb);
1565                         else
1566                                 lvb = req_capsule_server_swab_get(pill,
1567                                                 &RMF_DLM_LVB,
1568                                                 lustre_swab_ost_lvb);
1569                         if (unlikely(lvb == NULL)) {
1570                                 LDLM_ERROR(lock, "no LVB");
1571                                 RETURN(-EPROTO);
1572                         }
1573
1574                         memcpy(data, lvb, size);
1575                 } else if (size == sizeof(struct ost_lvb_v1)) {
1576                         struct ost_lvb *olvb = data;
1577
1578                         if (loc == RCL_CLIENT)
1579                                 lvb = req_capsule_client_swab_get(pill,
1580                                                 &RMF_DLM_LVB,
1581                                                 lustre_swab_ost_lvb_v1);
1582                         else
1583                                 lvb = req_capsule_server_sized_swab_get(pill,
1584                                                 &RMF_DLM_LVB, size,
1585                                                 lustre_swab_ost_lvb_v1);
1586                         if (unlikely(lvb == NULL)) {
1587                                 LDLM_ERROR(lock, "no LVB");
1588                                 RETURN(-EPROTO);
1589                         }
1590
1591                         memcpy(data, lvb, size);
1592                         olvb->lvb_mtime_ns = 0;
1593                         olvb->lvb_atime_ns = 0;
1594                         olvb->lvb_ctime_ns = 0;
1595                 } else {
1596                         LDLM_ERROR(lock, "Replied unexpected ost LVB size %d",
1597                                    size);
1598                         RETURN(-EINVAL);
1599                 }
1600                 break;
1601         case LVB_T_LQUOTA:
1602                 if (size == sizeof(struct lquota_lvb)) {
1603                         if (loc == RCL_CLIENT)
1604                                 lvb = req_capsule_client_swab_get(pill,
1605                                                 &RMF_DLM_LVB,
1606                                                 lustre_swab_lquota_lvb);
1607                         else
1608                                 lvb = req_capsule_server_swab_get(pill,
1609                                                 &RMF_DLM_LVB,
1610                                                 lustre_swab_lquota_lvb);
1611                         if (unlikely(lvb == NULL)) {
1612                                 LDLM_ERROR(lock, "no LVB");
1613                                 RETURN(-EPROTO);
1614                         }
1615
1616                         memcpy(data, lvb, size);
1617                 } else {
1618                         LDLM_ERROR(lock, "Replied unexpected lquota LVB size %d",
1619                                    size);
1620                         RETURN(-EINVAL);
1621                 }
1622                 break;
1623         case LVB_T_LAYOUT:
1624                 if (size == 0)
1625                         break;
1626
1627                 if (loc == RCL_CLIENT)
1628                         lvb = req_capsule_client_get(pill, &RMF_DLM_LVB);
1629                 else
1630                         lvb = req_capsule_server_get(pill, &RMF_DLM_LVB);
1631                 if (unlikely(lvb == NULL)) {
1632                         LDLM_ERROR(lock, "no LVB");
1633                         RETURN(-EPROTO);
1634                 }
1635
1636                 memcpy(data, lvb, size);
1637                 break;
1638         default:
1639                 LDLM_ERROR(lock, "Unknown LVB type: %d", lock->l_lvb_type);
1640                 libcfs_debug_dumpstack(NULL);
1641                 RETURN(-EINVAL);
1642         }
1643
1644         RETURN(0);
1645 }
1646
1647 /**
1648  * Create and fill in new LDLM lock with specified properties.
1649  * Returns a referenced lock
1650  */
1651 struct ldlm_lock *ldlm_lock_create(struct ldlm_namespace *ns,
1652                                    const struct ldlm_res_id *res_id,
1653                                    enum ldlm_type type,
1654                                    enum ldlm_mode mode,
1655                                    const struct ldlm_callback_suite *cbs,
1656                                    void *data, __u32 lvb_len,
1657                                    enum lvb_type lvb_type)
1658 {
1659         struct ldlm_lock        *lock;
1660         struct ldlm_resource    *res;
1661         int                     rc;
1662         ENTRY;
1663
1664         res = ldlm_resource_get(ns, NULL, res_id, type, 1);
1665         if (IS_ERR(res))
1666                 RETURN(ERR_CAST(res));
1667
1668         lock = ldlm_lock_new(res);
1669         if (lock == NULL)
1670                 RETURN(ERR_PTR(-ENOMEM));
1671
1672         lock->l_req_mode = mode;
1673         lock->l_ast_data = data;
1674         lock->l_pid = current_pid();
1675         if (ns_is_server(ns))
1676                 ldlm_set_ns_srv(lock);
1677         if (cbs) {
1678                 lock->l_blocking_ast = cbs->lcs_blocking;
1679                 lock->l_completion_ast = cbs->lcs_completion;
1680                 lock->l_glimpse_ast = cbs->lcs_glimpse;
1681         }
1682
1683         switch (type) {
1684         case LDLM_EXTENT:
1685                 rc = ldlm_extent_alloc_lock(lock);
1686                 break;
1687         case LDLM_IBITS:
1688                 rc = ldlm_inodebits_alloc_lock(lock);
1689                 break;
1690         default:
1691                 rc = 0;
1692         }
1693         if (rc)
1694                 GOTO(out, rc);
1695
1696         if (lvb_len) {
1697                 lock->l_lvb_len = lvb_len;
1698                 OBD_ALLOC_LARGE(lock->l_lvb_data, lvb_len);
1699                 if (lock->l_lvb_data == NULL)
1700                         GOTO(out, rc = -ENOMEM);
1701         }
1702
1703         lock->l_lvb_type = lvb_type;
1704         if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_NEW_LOCK))
1705                 GOTO(out, rc = -ENOENT);
1706
1707         RETURN(lock);
1708
1709 out:
1710         ldlm_lock_destroy(lock);
1711         LDLM_LOCK_RELEASE(lock);
1712         RETURN(ERR_PTR(rc));
1713 }
1714
1715 #ifdef HAVE_SERVER_SUPPORT
1716 static enum ldlm_error ldlm_lock_enqueue_helper(struct ldlm_lock *lock,
1717                                              __u64 *flags)
1718 {
1719         struct ldlm_resource *res = lock->l_resource;
1720         enum ldlm_error rc = ELDLM_OK;
1721         LIST_HEAD(rpc_list);
1722         ldlm_processing_policy policy;
1723
1724         ENTRY;
1725
1726         policy = ldlm_get_processing_policy(res);
1727 restart:
1728         policy(lock, flags, LDLM_PROCESS_ENQUEUE, &rc, &rpc_list);
1729         if (rc == ELDLM_OK && lock->l_granted_mode != lock->l_req_mode &&
1730             res->lr_type != LDLM_FLOCK) {
1731                 rc = ldlm_handle_conflict_lock(lock, flags, &rpc_list);
1732                 if (rc == -ERESTART)
1733                         GOTO(restart, rc);
1734         }
1735
1736         if (!list_empty(&rpc_list))
1737                 ldlm_discard_bl_list(&rpc_list);
1738
1739         RETURN(rc);
1740 }
1741 #endif
1742
1743 /**
1744  * Enqueue (request) a lock.
1745  *
1746  * Does not block. As a result of enqueue the lock would be put
1747  * into granted or waiting list.
1748  *
1749  * If namespace has intent policy sent and the lock has LDLM_FL_HAS_INTENT flag
1750  * set, skip all the enqueueing and delegate lock processing to intent policy
1751  * function.
1752  */
1753 enum ldlm_error ldlm_lock_enqueue(const struct lu_env *env,
1754                                   struct ldlm_namespace *ns,
1755                                   struct ldlm_lock **lockp,
1756                                   void *cookie, __u64 *flags)
1757 {
1758         struct ldlm_lock *lock = *lockp;
1759         struct ldlm_resource *res = lock->l_resource;
1760         int local = ns_is_client(ldlm_res_to_ns(res));
1761         enum ldlm_error rc = ELDLM_OK;
1762         struct ldlm_interval *node = NULL;
1763 #ifdef HAVE_SERVER_SUPPORT
1764         bool reconstruct = false;
1765 #endif
1766         ENTRY;
1767
1768         /* policies are not executed on the client or during replay */
1769         if ((*flags & (LDLM_FL_HAS_INTENT|LDLM_FL_REPLAY)) == LDLM_FL_HAS_INTENT
1770             && !local && ns->ns_policy) {
1771                 rc = ns->ns_policy(env, ns, lockp, cookie, lock->l_req_mode,
1772                                    *flags, NULL);
1773                 if (rc == ELDLM_LOCK_REPLACED) {
1774                         /* The lock that was returned has already been granted,
1775                          * and placed into lockp.  If it's not the same as the
1776                          * one we passed in, then destroy the old one and our
1777                          * work here is done. */
1778                         if (lock != *lockp) {
1779                                 ldlm_lock_destroy(lock);
1780                                 LDLM_LOCK_RELEASE(lock);
1781                         }
1782                         *flags |= LDLM_FL_LOCK_CHANGED;
1783                         RETURN(0);
1784                 } else if (rc != ELDLM_OK &&
1785                            ldlm_is_granted(lock)) {
1786                         LASSERT(*flags & LDLM_FL_RESENT);
1787                         /* It may happen that ns_policy returns an error in
1788                          * resend case, object may be unlinked or just some
1789                          * error occurs. It is unclear if lock reached the
1790                          * client in the original reply, just leave the lock on
1791                          * server, not returning it again to client. Due to
1792                          * LU-6529, the server will not OOM. */
1793                         RETURN(rc);
1794                 } else if (rc != ELDLM_OK ||
1795                            (rc == ELDLM_OK && (*flags & LDLM_FL_INTENT_ONLY))) {
1796                         ldlm_lock_destroy(lock);
1797                         RETURN(rc);
1798                 }
1799         }
1800
1801         if (*flags & LDLM_FL_RESENT) {
1802                 /* Reconstruct LDLM_FL_SRV_ENQ_MASK @flags for reply.
1803                  * Set LOCK_CHANGED always.
1804                  * Check if the lock is granted for BLOCK_GRANTED.
1805                  * Take NO_TIMEOUT from the lock as it is inherited through
1806                  * LDLM_FL_INHERIT_MASK */
1807                 *flags |= LDLM_FL_LOCK_CHANGED;
1808                 if (!ldlm_is_granted(lock))
1809                         *flags |= LDLM_FL_BLOCK_GRANTED;
1810                 *flags |= lock->l_flags & LDLM_FL_NO_TIMEOUT;
1811                 RETURN(ELDLM_OK);
1812         }
1813
1814         /* For a replaying lock, it might be already in granted list. So
1815          * unlinking the lock will cause the interval node to be freed, we
1816          * have to allocate the interval node early otherwise we can't regrant
1817          * this lock in the future. - jay */
1818         if (!local && (*flags & LDLM_FL_REPLAY) && res->lr_type == LDLM_EXTENT)
1819                 OBD_SLAB_ALLOC_PTR_GFP(node, ldlm_interval_slab, GFP_NOFS);
1820
1821 #ifdef HAVE_SERVER_SUPPORT
1822         reconstruct = !local && res->lr_type == LDLM_FLOCK &&
1823                       !(*flags & LDLM_FL_TEST_LOCK);
1824         if (reconstruct) {
1825                 rc = req_can_reconstruct(cookie, NULL);
1826                 if (rc != 0) {
1827                         if (rc == 1)
1828                                 rc = 0;
1829                         RETURN(rc);
1830                 }
1831         }
1832 #endif
1833
1834         lock_res_and_lock(lock);
1835         if (local && ldlm_is_granted(lock)) {
1836                 /* The server returned a blocked lock, but it was granted
1837                  * before we got a chance to actually enqueue it.  We don't
1838                  * need to do anything else. */
1839                 *flags &= ~LDLM_FL_BLOCKED_MASK;
1840                 GOTO(out, rc = ELDLM_OK);
1841         }
1842
1843         ldlm_resource_unlink_lock(lock);
1844         if (res->lr_type == LDLM_EXTENT && lock->l_tree_node == NULL) {
1845                 if (node == NULL) {
1846                         ldlm_lock_destroy_nolock(lock);
1847                         GOTO(out, rc = -ENOMEM);
1848                 }
1849
1850                 INIT_LIST_HEAD(&node->li_group);
1851                 ldlm_interval_attach(node, lock);
1852                 node = NULL;
1853         }
1854
1855         /* Some flags from the enqueue want to make it into the AST, via the
1856          * lock's l_flags. */
1857         if (*flags & LDLM_FL_AST_DISCARD_DATA)
1858                 ldlm_set_ast_discard_data(lock);
1859         if (*flags & LDLM_FL_TEST_LOCK)
1860                 ldlm_set_test_lock(lock);
1861         if (*flags & LDLM_FL_COS_INCOMPAT)
1862                 ldlm_set_cos_incompat(lock);
1863         if (*flags & LDLM_FL_COS_ENABLED)
1864                 ldlm_set_cos_enabled(lock);
1865
1866         /* This distinction between local lock trees is very important; a client
1867          * namespace only has information about locks taken by that client, and
1868          * thus doesn't have enough information to decide for itself if it can
1869          * be granted (below).  In this case, we do exactly what the server
1870          * tells us to do, as dictated by the 'flags'.
1871          *
1872          * We do exactly the same thing during recovery, when the server is
1873          * more or less trusting the clients not to lie.
1874          *
1875          * FIXME (bug 268): Detect obvious lies by checking compatibility in
1876          * granted queue. */
1877         if (local) {
1878                 if (*flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED))
1879                         ldlm_resource_add_lock(res, &res->lr_waiting, lock);
1880                 else
1881                         ldlm_grant_lock(lock, NULL);
1882                 GOTO(out, rc = ELDLM_OK);
1883 #ifdef HAVE_SERVER_SUPPORT
1884         } else if (*flags & LDLM_FL_REPLAY) {
1885                 if (*flags & LDLM_FL_BLOCK_WAIT) {
1886                         ldlm_resource_add_lock(res, &res->lr_waiting, lock);
1887                         GOTO(out, rc = ELDLM_OK);
1888                 } else if (*flags & LDLM_FL_BLOCK_GRANTED) {
1889                         ldlm_grant_lock(lock, NULL);
1890                         GOTO(out, rc = ELDLM_OK);
1891                 }
1892                 /* If no flags, fall through to normal enqueue path. */
1893         }
1894
1895         rc = ldlm_lock_enqueue_helper(lock, flags);
1896         GOTO(out, rc);
1897 #else
1898         } else {
1899                 CERROR("This is client-side-only module, cannot handle "
1900                        "LDLM_NAMESPACE_SERVER resource type lock.\n");
1901                 LBUG();
1902         }
1903 #endif
1904
1905 out:
1906         unlock_res_and_lock(lock);
1907
1908 #ifdef HAVE_SERVER_SUPPORT
1909         if (reconstruct) {
1910                 struct ptlrpc_request *req = cookie;
1911
1912                 tgt_mk_reply_data(NULL, NULL,
1913                                   &req->rq_export->exp_target_data,
1914                                   req, 0, NULL, false, 0);
1915         }
1916 #endif
1917         if (node)
1918                 OBD_SLAB_FREE(node, ldlm_interval_slab, sizeof(*node));
1919         return rc;
1920 }
1921
1922 #ifdef HAVE_SERVER_SUPPORT
1923 /**
1924  * Iterate through all waiting locks on a given resource queue and attempt to
1925  * grant them.
1926  *
1927  * Must be called with resource lock held.
1928  */
1929 int ldlm_reprocess_queue(struct ldlm_resource *res, struct list_head *queue,
1930                          struct list_head *work_list,
1931                          enum ldlm_process_intention intention,
1932                          struct ldlm_lock *hint)
1933 {
1934         struct list_head *tmp, *pos;
1935         ldlm_processing_policy policy;
1936         __u64 flags;
1937         int rc = LDLM_ITER_CONTINUE;
1938         enum ldlm_error err;
1939         LIST_HEAD(bl_ast_list);
1940
1941         ENTRY;
1942
1943         check_res_locked(res);
1944
1945         policy = ldlm_get_processing_policy(res);
1946         LASSERT(policy);
1947         LASSERT(intention == LDLM_PROCESS_RESCAN ||
1948                 intention == LDLM_PROCESS_RECOVERY);
1949
1950 restart:
1951         list_for_each_safe(tmp, pos, queue) {
1952                 struct ldlm_lock *pending;
1953                 LIST_HEAD(rpc_list);
1954
1955                 pending = list_entry(tmp, struct ldlm_lock, l_res_link);
1956
1957                 CDEBUG(D_INFO, "Reprocessing lock %p\n", pending);
1958
1959                 flags = 0;
1960                 rc = policy(pending, &flags, intention, &err, &rpc_list);
1961                 if (pending->l_granted_mode == pending->l_req_mode ||
1962                     res->lr_type == LDLM_FLOCK) {
1963                         list_splice(&rpc_list, work_list);
1964                 } else {
1965                         list_splice(&rpc_list, &bl_ast_list);
1966                 }
1967                 /*
1968                  * When this is called from recovery done, we always want
1969                  * to scan the whole list no matter what 'rc' is returned.
1970                  */
1971                 if (rc != LDLM_ITER_CONTINUE &&
1972                     intention == LDLM_PROCESS_RESCAN)
1973                         break;
1974         }
1975
1976         if (!list_empty(&bl_ast_list)) {
1977                 unlock_res(res);
1978
1979                 rc = ldlm_run_ast_work(ldlm_res_to_ns(res), &bl_ast_list,
1980                                        LDLM_WORK_BL_AST);
1981
1982                 lock_res(res);
1983                 if (rc == -ERESTART)
1984                         GOTO(restart, rc);
1985         }
1986
1987         if (!list_empty(&bl_ast_list))
1988                 ldlm_discard_bl_list(&bl_ast_list);
1989
1990         RETURN(intention == LDLM_PROCESS_RESCAN ? rc : LDLM_ITER_CONTINUE);
1991 }
1992
1993 /**
1994  * Conflicting locks are detected for a lock to be enqueued, add the lock
1995  * into waiting list and send blocking ASTs to the conflicting locks.
1996  *
1997  * \param[in] lock              The lock to be enqueued.
1998  * \param[out] flags            Lock flags for the lock to be enqueued.
1999  * \param[in] rpc_list          Conflicting locks list.
2000  *
2001  * \retval -ERESTART:   Some lock was instantly canceled while sending
2002  *                      blocking ASTs, caller needs to re-check conflicting
2003  *                      locks.
2004  * \retval -EAGAIN:     Lock was destroyed, caller should return error.
2005  * \reval 0:            Lock is successfully added in waiting list.
2006  */
2007 int ldlm_handle_conflict_lock(struct ldlm_lock *lock, __u64 *flags,
2008                               struct list_head *rpc_list)
2009 {
2010         struct ldlm_resource *res = lock->l_resource;
2011         int rc;
2012         ENTRY;
2013
2014         check_res_locked(res);
2015
2016         /* If either of the compat_queue()s returned failure, then we
2017          * have ASTs to send and must go onto the waiting list.
2018          *
2019          * bug 2322: we used to unlink and re-add here, which was a
2020          * terrible folly -- if we goto restart, we could get
2021          * re-ordered!  Causes deadlock, because ASTs aren't sent! */
2022         if (list_empty(&lock->l_res_link))
2023                 ldlm_resource_add_lock(res, &res->lr_waiting, lock);
2024         unlock_res(res);
2025
2026         rc = ldlm_run_ast_work(ldlm_res_to_ns(res), rpc_list,
2027                                LDLM_WORK_BL_AST);
2028
2029         if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_OST_FAIL_RACE) &&
2030             !ns_is_client(ldlm_res_to_ns(res)))
2031                 class_fail_export(lock->l_export);
2032
2033         lock_res(res);
2034         if (rc == -ERESTART) {
2035                 /* 15715: The lock was granted and destroyed after
2036                  * resource lock was dropped. Interval node was freed
2037                  * in ldlm_lock_destroy. Anyway, this always happens
2038                  * when a client is being evicted. So it would be
2039                  * ok to return an error. -jay */
2040                 if (ldlm_is_destroyed(lock))
2041                         RETURN(-EAGAIN);
2042
2043                 /* lock was granted while resource was unlocked. */
2044                 if (ldlm_is_granted(lock)) {
2045                         /* bug 11300: if the lock has been granted,
2046                          * break earlier because otherwise, we will go
2047                          * to restart and ldlm_resource_unlink will be
2048                          * called and it causes the interval node to be
2049                          * freed. Then we will fail at
2050                          * ldlm_extent_add_lock() */
2051                         *flags &= ~LDLM_FL_BLOCKED_MASK;
2052                         RETURN(0);
2053                 }
2054
2055                 RETURN(rc);
2056         }
2057         *flags |= LDLM_FL_BLOCK_GRANTED;
2058
2059         RETURN(0);
2060 }
2061
2062 /**
2063  * Discard all AST work items from list.
2064  *
2065  * If for whatever reason we do not want to send ASTs to conflicting locks
2066  * anymore, disassemble the list with this function.
2067  */
2068 void ldlm_discard_bl_list(struct list_head *bl_list)
2069 {
2070         struct ldlm_lock *lock, *tmp;
2071
2072         ENTRY;
2073
2074         list_for_each_entry_safe(lock, tmp, bl_list, l_bl_ast) {
2075                 LASSERT(!list_empty(&lock->l_bl_ast));
2076                 list_del_init(&lock->l_bl_ast);
2077                 ldlm_clear_ast_sent(lock);
2078                 LASSERT(lock->l_bl_ast_run == 0);
2079                 ldlm_clear_blocking_lock(lock);
2080                 LDLM_LOCK_RELEASE(lock);
2081         }
2082         EXIT;
2083 }
2084
2085 /**
2086  * Process a call to blocking AST callback for a lock in ast_work list
2087  */
2088 static int
2089 ldlm_work_bl_ast_lock(struct ptlrpc_request_set *rqset, void *opaq)
2090 {
2091         struct ldlm_cb_set_arg *arg = opaq;
2092         struct ldlm_lock *lock;
2093         struct ldlm_lock_desc d;
2094         struct ldlm_bl_desc bld;
2095         int rc;
2096
2097         ENTRY;
2098
2099         if (list_empty(arg->list))
2100                 RETURN(-ENOENT);
2101
2102         lock = list_entry(arg->list->next, struct ldlm_lock, l_bl_ast);
2103
2104         /* nobody should touch l_bl_ast but some locks in the list may become
2105          * granted after lock convert or COS downgrade, these locks should be
2106          * just skipped here and removed from the list.
2107          */
2108         lock_res_and_lock(lock);
2109         list_del_init(&lock->l_bl_ast);
2110
2111         /* lock is not blocking lock anymore, but was kept in the list because
2112          * it can managed only here.
2113          */
2114         if (!ldlm_is_ast_sent(lock)) {
2115                 unlock_res_and_lock(lock);
2116                 LDLM_LOCK_RELEASE(lock);
2117                 RETURN(0);
2118         }
2119
2120         LASSERT(lock->l_blocking_lock);
2121         ldlm_lock2desc(lock->l_blocking_lock, &d);
2122         /* copy blocking lock ibits in cancel_bits as well,
2123          * new client may use them for lock convert and it is
2124          * important to use new field to convert locks from
2125          * new servers only
2126          */
2127         d.l_policy_data.l_inodebits.cancel_bits =
2128                 lock->l_blocking_lock->l_policy_data.l_inodebits.bits;
2129
2130         /* Blocking lock is being destroyed here but some information about it
2131          * may be needed inside l_blocking_ast() function below,
2132          * e.g. in mdt_blocking_ast(). So save needed data in bl_desc.
2133          */
2134         bld.bl_same_client = lock->l_client_cookie ==
2135                              lock->l_blocking_lock->l_client_cookie;
2136         bld.bl_cos_incompat = ldlm_is_cos_incompat(lock->l_blocking_lock);
2137         arg->bl_desc = &bld;
2138
2139         LASSERT(ldlm_is_ast_sent(lock));
2140         LASSERT(lock->l_bl_ast_run == 0);
2141         lock->l_bl_ast_run++;
2142         ldlm_clear_blocking_lock(lock);
2143         unlock_res_and_lock(lock);
2144
2145         rc = lock->l_blocking_ast(lock, &d, (void *)arg, LDLM_CB_BLOCKING);
2146
2147         LDLM_LOCK_RELEASE(lock);
2148
2149         RETURN(rc);
2150 }
2151
2152 /**
2153  * Process a call to revocation AST callback for a lock in ast_work list
2154  */
2155 static int
2156 ldlm_work_revoke_ast_lock(struct ptlrpc_request_set *rqset, void *opaq)
2157 {
2158         struct ldlm_cb_set_arg *arg = opaq;
2159         struct ldlm_lock_desc   desc;
2160         int                     rc;
2161         struct ldlm_lock       *lock;
2162         ENTRY;
2163
2164         if (list_empty(arg->list))
2165                 RETURN(-ENOENT);
2166
2167         lock = list_entry(arg->list->next, struct ldlm_lock, l_rk_ast);
2168         list_del_init(&lock->l_rk_ast);
2169
2170         /* the desc just pretend to exclusive */
2171         ldlm_lock2desc(lock, &desc);
2172         desc.l_req_mode = LCK_EX;
2173         desc.l_granted_mode = 0;
2174
2175         rc = lock->l_blocking_ast(lock, &desc, (void*)arg, LDLM_CB_BLOCKING);
2176         LDLM_LOCK_RELEASE(lock);
2177
2178         RETURN(rc);
2179 }
2180
2181 /**
2182  * Process a call to glimpse AST callback for a lock in ast_work list
2183  */
2184 int ldlm_work_gl_ast_lock(struct ptlrpc_request_set *rqset, void *opaq)
2185 {
2186         struct ldlm_cb_set_arg          *arg = opaq;
2187         struct ldlm_glimpse_work        *gl_work;
2188         struct ldlm_lock                *lock;
2189         int                              rc = 0;
2190         ENTRY;
2191
2192         if (list_empty(arg->list))
2193                 RETURN(-ENOENT);
2194
2195         gl_work = list_entry(arg->list->next, struct ldlm_glimpse_work,
2196                                  gl_list);
2197         list_del_init(&gl_work->gl_list);
2198
2199         lock = gl_work->gl_lock;
2200
2201         /* transfer the glimpse descriptor to ldlm_cb_set_arg */
2202         arg->gl_desc = gl_work->gl_desc;
2203         arg->gl_interpret_reply = gl_work->gl_interpret_reply;
2204         arg->gl_interpret_data = gl_work->gl_interpret_data;
2205
2206         /* invoke the actual glimpse callback */
2207         if (lock->l_glimpse_ast(lock, (void*)arg) == 0)
2208                 rc = 1;
2209
2210         LDLM_LOCK_RELEASE(lock);
2211         if (gl_work->gl_flags & LDLM_GL_WORK_SLAB_ALLOCATED)
2212                 OBD_SLAB_FREE_PTR(gl_work, ldlm_glimpse_work_kmem);
2213         else
2214                 OBD_FREE_PTR(gl_work);
2215
2216         RETURN(rc);
2217 }
2218 #endif
2219
2220 /**
2221  * Process a call to completion AST callback for a lock in ast_work list
2222  */
2223 static int
2224 ldlm_work_cp_ast_lock(struct ptlrpc_request_set *rqset, void *opaq)
2225 {
2226         struct ldlm_cb_set_arg *arg = opaq;
2227         struct ldlm_lock *lock;
2228         ldlm_completion_callback completion_callback;
2229         int rc = 0;
2230
2231         ENTRY;
2232
2233         if (list_empty(arg->list))
2234                 RETURN(-ENOENT);
2235
2236         lock = list_entry(arg->list->next, struct ldlm_lock, l_cp_ast);
2237
2238         /* It's possible to receive a completion AST before we've set
2239          * the l_completion_ast pointer: either because the AST arrived
2240          * before the reply, or simply because there's a small race
2241          * window between receiving the reply and finishing the local
2242          * enqueue. (bug 842)
2243          *
2244          * This can't happen with the blocking_ast, however, because we
2245          * will never call the local blocking_ast until we drop our
2246          * reader/writer reference, which we won't do until we get the
2247          * reply and finish enqueueing. */
2248
2249         /* nobody should touch l_cp_ast */
2250         lock_res_and_lock(lock);
2251         list_del_init(&lock->l_cp_ast);
2252         LASSERT(ldlm_is_cp_reqd(lock));
2253         /* save l_completion_ast since it can be changed by
2254          * mds_intent_policy(), see bug 14225 */
2255         completion_callback = lock->l_completion_ast;
2256         ldlm_clear_cp_reqd(lock);
2257         unlock_res_and_lock(lock);
2258
2259         if (completion_callback != NULL)
2260                 rc = completion_callback(lock, 0, (void *)arg);
2261         LDLM_LOCK_RELEASE(lock);
2262
2263         RETURN(rc);
2264 }
2265
2266 /**
2267  * Process list of locks in need of ASTs being sent.
2268  *
2269  * Used on server to send multiple ASTs together instead of sending one by
2270  * one.
2271  */
2272 int ldlm_run_ast_work(struct ldlm_namespace *ns, struct list_head *rpc_list,
2273                       ldlm_desc_ast_t ast_type)
2274 {
2275         struct ldlm_cb_set_arg *arg;
2276         set_producer_func work_ast_lock;
2277         int rc;
2278
2279         if (list_empty(rpc_list))
2280                 RETURN(0);
2281
2282         OBD_ALLOC_PTR(arg);
2283         if (arg == NULL)
2284                 RETURN(-ENOMEM);
2285
2286         atomic_set(&arg->restart, 0);
2287         arg->list = rpc_list;
2288
2289         switch (ast_type) {
2290         case LDLM_WORK_CP_AST:
2291                 arg->type = LDLM_CP_CALLBACK;
2292                 work_ast_lock = ldlm_work_cp_ast_lock;
2293                 break;
2294 #ifdef HAVE_SERVER_SUPPORT
2295         case LDLM_WORK_BL_AST:
2296                 arg->type = LDLM_BL_CALLBACK;
2297                 work_ast_lock = ldlm_work_bl_ast_lock;
2298                 break;
2299         case LDLM_WORK_REVOKE_AST:
2300                 arg->type = LDLM_BL_CALLBACK;
2301                 work_ast_lock = ldlm_work_revoke_ast_lock;
2302                 break;
2303         case LDLM_WORK_GL_AST:
2304                 arg->type = LDLM_GL_CALLBACK;
2305                 work_ast_lock = ldlm_work_gl_ast_lock;
2306                 break;
2307 #endif
2308         default:
2309                 LBUG();
2310         }
2311
2312         /* We create a ptlrpc request set with flow control extension.
2313          * This request set will use the work_ast_lock function to produce new
2314          * requests and will send a new request each time one completes in order
2315          * to keep the number of requests in flight to ns_max_parallel_ast */
2316         arg->set = ptlrpc_prep_fcset(ns->ns_max_parallel_ast ? : UINT_MAX,
2317                                      work_ast_lock, arg);
2318         if (arg->set == NULL)
2319                 GOTO(out, rc = -ENOMEM);
2320
2321         ptlrpc_set_wait(NULL, arg->set);
2322         ptlrpc_set_destroy(arg->set);
2323
2324         rc = atomic_read(&arg->restart) ? -ERESTART : 0;
2325         GOTO(out, rc);
2326 out:
2327         OBD_FREE_PTR(arg);
2328         return rc;
2329 }
2330
2331 /**
2332  * Try to grant all waiting locks on a resource.
2333  *
2334  * Calls ldlm_reprocess_queue on waiting queue.
2335  *
2336  * Typically called after some resource locks are cancelled to see
2337  * if anything could be granted as a result of the cancellation.
2338  */
2339 static void __ldlm_reprocess_all(struct ldlm_resource *res,
2340                                  enum ldlm_process_intention intention,
2341                                  struct ldlm_lock *hint)
2342 {
2343         LIST_HEAD(rpc_list);
2344 #ifdef HAVE_SERVER_SUPPORT
2345         ldlm_reprocessing_policy reprocess;
2346         struct obd_device *obd;
2347         int rc;
2348
2349         ENTRY;
2350
2351         /* Local lock trees don't get reprocessed. */
2352         if (ns_is_client(ldlm_res_to_ns(res))) {
2353                 EXIT;
2354                 return;
2355         }
2356
2357         /* Disable reprocess during lock replay stage but allow during
2358          * request replay stage.
2359          */
2360         obd = ldlm_res_to_ns(res)->ns_obd;
2361         if (obd->obd_recovering &&
2362             atomic_read(&obd->obd_req_replay_clients) == 0)
2363                 RETURN_EXIT;
2364 restart:
2365         lock_res(res);
2366         reprocess = ldlm_get_reprocessing_policy(res);
2367         reprocess(res, &res->lr_waiting, &rpc_list, intention, hint);
2368         unlock_res(res);
2369
2370         rc = ldlm_run_ast_work(ldlm_res_to_ns(res), &rpc_list,
2371                                LDLM_WORK_CP_AST);
2372         if (rc == -ERESTART) {
2373                 LASSERT(list_empty(&rpc_list));
2374                 goto restart;
2375         }
2376 #else
2377         ENTRY;
2378
2379         if (!ns_is_client(ldlm_res_to_ns(res))) {
2380                 CERROR("This is client-side-only module, cannot handle "
2381                        "LDLM_NAMESPACE_SERVER resource type lock.\n");
2382                 LBUG();
2383         }
2384 #endif
2385         EXIT;
2386 }
2387
2388 void ldlm_reprocess_all(struct ldlm_resource *res, struct ldlm_lock *hint)
2389 {
2390         __ldlm_reprocess_all(res, LDLM_PROCESS_RESCAN, hint);
2391 }
2392 EXPORT_SYMBOL(ldlm_reprocess_all);
2393
2394 static int ldlm_reprocess_res(struct cfs_hash *hs, struct cfs_hash_bd *bd,
2395                               struct hlist_node *hnode, void *arg)
2396 {
2397         struct ldlm_resource *res = cfs_hash_object(hs, hnode);
2398
2399         /* This is only called once after recovery done. LU-8306. */
2400         __ldlm_reprocess_all(res, LDLM_PROCESS_RECOVERY, NULL);
2401         return 0;
2402 }
2403
2404 /**
2405  * Iterate through all resources on a namespace attempting to grant waiting
2406  * locks.
2407  */
2408 void ldlm_reprocess_recovery_done(struct ldlm_namespace *ns)
2409 {
2410         ENTRY;
2411
2412         if (ns != NULL) {
2413                 cfs_hash_for_each_nolock(ns->ns_rs_hash,
2414                                          ldlm_reprocess_res, NULL, 0);
2415         }
2416         EXIT;
2417 }
2418
2419 /**
2420  * Helper function to call blocking AST for LDLM lock \a lock in a
2421  * "cancelling" mode.
2422  */
2423 void ldlm_cancel_callback(struct ldlm_lock *lock)
2424 {
2425         check_res_locked(lock->l_resource);
2426         if (!ldlm_is_cancel(lock)) {
2427                 ldlm_set_cancel(lock);
2428                 if (lock->l_blocking_ast) {
2429                         unlock_res_and_lock(lock);
2430                         lock->l_blocking_ast(lock, NULL, lock->l_ast_data,
2431                                              LDLM_CB_CANCELING);
2432                         lock_res_and_lock(lock);
2433                 } else {
2434                         LDLM_DEBUG(lock, "no blocking ast");
2435                 }
2436
2437                 /* only canceller can set bl_done bit */
2438                 ldlm_set_bl_done(lock);
2439                 wake_up_all(&lock->l_waitq);
2440         } else if (!ldlm_is_bl_done(lock)) {
2441                 /* The lock is guaranteed to have been canceled once
2442                  * returning from this function. */
2443                 unlock_res_and_lock(lock);
2444                 wait_event_idle(lock->l_waitq, is_bl_done(lock));
2445                 lock_res_and_lock(lock);
2446         }
2447 }
2448
2449 /**
2450  * Remove skiplist-enabled LDLM lock \a req from granted list
2451  */
2452 void ldlm_unlink_lock_skiplist(struct ldlm_lock *req)
2453 {
2454         if (req->l_resource->lr_type != LDLM_PLAIN &&
2455             req->l_resource->lr_type != LDLM_IBITS)
2456                 return;
2457
2458         list_del_init(&req->l_sl_policy);
2459         list_del_init(&req->l_sl_mode);
2460 }
2461
2462 /**
2463  * Attempts to cancel LDLM lock \a lock that has no reader/writer references.
2464  */
2465 void ldlm_lock_cancel(struct ldlm_lock *lock)
2466 {
2467         struct ldlm_resource *res;
2468         struct ldlm_namespace *ns;
2469         ENTRY;
2470
2471         lock_res_and_lock(lock);
2472
2473         res = lock->l_resource;
2474         ns  = ldlm_res_to_ns(res);
2475
2476         /* Please do not, no matter how tempting, remove this LBUG without
2477          * talking to me first. -phik */
2478         if (lock->l_readers || lock->l_writers) {
2479                 LDLM_ERROR(lock, "lock still has references");
2480                 unlock_res_and_lock(lock);
2481                 LBUG();
2482         }
2483
2484         if (ldlm_is_waited(lock))
2485                 ldlm_del_waiting_lock(lock);
2486
2487         /* Releases cancel callback. */
2488         ldlm_cancel_callback(lock);
2489
2490         /* Yes, second time, just in case it was added again while we were
2491          * running with no res lock in ldlm_cancel_callback */
2492         if (ldlm_is_waited(lock))
2493                 ldlm_del_waiting_lock(lock);
2494
2495         ldlm_resource_unlink_lock(lock);
2496         ldlm_lock_destroy_nolock(lock);
2497
2498         if (ldlm_is_granted(lock))
2499                 ldlm_pool_del(&ns->ns_pool, lock);
2500
2501         /* Make sure we will not be called again for same lock what is possible
2502          * if not to zero out lock->l_granted_mode */
2503         lock->l_granted_mode = LCK_MINMODE;
2504         unlock_res_and_lock(lock);
2505
2506         EXIT;
2507 }
2508 EXPORT_SYMBOL(ldlm_lock_cancel);
2509
2510 /**
2511  * Set opaque data into the lock that only makes sense to upper layer.
2512  */
2513 int ldlm_lock_set_data(const struct lustre_handle *lockh, void *data)
2514 {
2515         struct ldlm_lock *lock = ldlm_handle2lock(lockh);
2516         int rc = -EINVAL;
2517         ENTRY;
2518
2519         if (lock) {
2520                 if (lock->l_ast_data == NULL)
2521                         lock->l_ast_data = data;
2522                 if (lock->l_ast_data == data)
2523                         rc = 0;
2524                 LDLM_LOCK_PUT(lock);
2525         }
2526         RETURN(rc);
2527 }
2528 EXPORT_SYMBOL(ldlm_lock_set_data);
2529
2530 struct export_cl_data {
2531         const struct lu_env     *ecl_env;
2532         struct obd_export       *ecl_exp;
2533         int                     ecl_loop;
2534 };
2535
2536 static void ldlm_cancel_lock_for_export(struct obd_export *exp,
2537                                         struct ldlm_lock *lock,
2538                                         struct export_cl_data *ecl)
2539 {
2540         struct ldlm_resource *res;
2541
2542         res = ldlm_resource_getref(lock->l_resource);
2543
2544         ldlm_lvbo_update(res, lock, NULL, 1);
2545         ldlm_lock_cancel(lock);
2546         if (!exp->exp_obd->obd_stopping)
2547                 ldlm_reprocess_all(res, lock);
2548         ldlm_resource_putref(res);
2549
2550         ecl->ecl_loop++;
2551         if ((ecl->ecl_loop & -ecl->ecl_loop) == ecl->ecl_loop) {
2552                 CDEBUG(D_INFO, "Export %p, %d locks cancelled.\n",
2553                        exp, ecl->ecl_loop);
2554         }
2555 }
2556
2557 /**
2558  * Iterator function for ldlm_export_cancel_locks.
2559  * Cancels passed locks.
2560  */
2561 static int
2562 ldlm_cancel_locks_for_export_cb(struct cfs_hash *hs, struct cfs_hash_bd *bd,
2563                                 struct hlist_node *hnode, void *data)
2564
2565 {
2566         struct export_cl_data   *ecl = (struct export_cl_data *)data;
2567         struct obd_export       *exp  = ecl->ecl_exp;
2568         struct ldlm_lock        *lock = cfs_hash_object(hs, hnode);
2569
2570         LDLM_LOCK_GET(lock);
2571         ldlm_cancel_lock_for_export(exp, lock, ecl);
2572         LDLM_LOCK_RELEASE(lock);
2573
2574         return 0;
2575 }
2576
2577 /**
2578  * Cancel all blocked locks for given export.
2579  *
2580  * Typically called on client disconnection/eviction
2581  */
2582 int ldlm_export_cancel_blocked_locks(struct obd_export *exp)
2583 {
2584         struct lu_env env;
2585         struct export_cl_data   ecl = {
2586                 .ecl_exp        = exp,
2587                 .ecl_loop       = 0,
2588         };
2589         int rc;
2590
2591         rc = lu_env_init(&env, LCT_DT_THREAD);
2592         if (rc)
2593                 RETURN(rc);
2594         ecl.ecl_env = &env;
2595
2596         while (!list_empty(&exp->exp_bl_list)) {
2597                 struct ldlm_lock *lock;
2598
2599                 spin_lock_bh(&exp->exp_bl_list_lock);
2600                 if (!list_empty(&exp->exp_bl_list)) {
2601                         lock = list_entry(exp->exp_bl_list.next,
2602                                           struct ldlm_lock, l_exp_list);
2603                         LDLM_LOCK_GET(lock);
2604                         list_del_init(&lock->l_exp_list);
2605                 } else {
2606                         lock = NULL;
2607                 }
2608                 spin_unlock_bh(&exp->exp_bl_list_lock);
2609
2610                 if (lock == NULL)
2611                         break;
2612
2613                 ldlm_cancel_lock_for_export(exp, lock, &ecl);
2614                 LDLM_LOCK_RELEASE(lock);
2615         }
2616
2617         lu_env_fini(&env);
2618
2619         CDEBUG(D_DLMTRACE, "Export %p, canceled %d locks, "
2620                "left on hash table %d.\n", exp, ecl.ecl_loop,
2621                atomic_read(&exp->exp_lock_hash->hs_count));
2622
2623         return ecl.ecl_loop;
2624 }
2625
2626 /**
2627  * Cancel all locks for given export.
2628  *
2629  * Typically called after client disconnection/eviction
2630  */
2631 int ldlm_export_cancel_locks(struct obd_export *exp)
2632 {
2633         struct export_cl_data ecl;
2634         struct lu_env env;
2635         int rc;
2636
2637         rc = lu_env_init(&env, LCT_DT_THREAD);
2638         if (rc)
2639                 RETURN(rc);
2640         ecl.ecl_env = &env;
2641         ecl.ecl_exp = exp;
2642         ecl.ecl_loop = 0;
2643
2644         cfs_hash_for_each_empty(exp->exp_lock_hash,
2645                                 ldlm_cancel_locks_for_export_cb, &ecl);
2646
2647         CDEBUG(D_DLMTRACE, "Export %p, canceled %d locks, "
2648                "left on hash table %d.\n", exp, ecl.ecl_loop,
2649                atomic_read(&exp->exp_lock_hash->hs_count));
2650
2651         if (ecl.ecl_loop > 0 &&
2652             atomic_read(&exp->exp_lock_hash->hs_count) == 0 &&
2653             exp->exp_obd->obd_stopping)
2654                 ldlm_reprocess_recovery_done(exp->exp_obd->obd_namespace);
2655
2656         lu_env_fini(&env);
2657
2658         return ecl.ecl_loop;
2659 }
2660
2661 /**
2662  * Downgrade an PW/EX lock to COS | CR mode.
2663  *
2664  * A lock mode convertion from PW/EX mode to less conflict mode. The
2665  * convertion may fail if lock was canceled before downgrade, but it doesn't
2666  * indicate any problem, because such lock has no reader or writer, and will
2667  * be released soon.
2668  *
2669  * Used by Commit on Sharing (COS) code to force object changes commit in case
2670  * of conflict. Converted lock is considered as new lock and all blocking AST
2671  * things are cleared, so any pending or new blocked lock on that lock will
2672  * cause new call to blocking_ast and force resource object commit.
2673  *
2674  * Also used by layout_change to replace EX lock to CR lock.
2675  *
2676  * \param lock A lock to convert
2677  * \param new_mode new lock mode
2678  */
2679 void ldlm_lock_mode_downgrade(struct ldlm_lock *lock, enum ldlm_mode new_mode)
2680 {
2681 #ifdef HAVE_SERVER_SUPPORT
2682         ENTRY;
2683
2684         LASSERT(new_mode == LCK_COS || new_mode == LCK_CR);
2685
2686         lock_res_and_lock(lock);
2687
2688         if (!(lock->l_granted_mode & (LCK_PW | LCK_EX))) {
2689                 unlock_res_and_lock(lock);
2690
2691                 LASSERT(lock->l_granted_mode == LCK_MINMODE);
2692                 LDLM_DEBUG(lock, "lock was canceled before downgrade");
2693                 RETURN_EXIT;
2694         }
2695
2696         ldlm_resource_unlink_lock(lock);
2697         /*
2698          * Remove the lock from pool as it will be added again in
2699          * ldlm_grant_lock() called below.
2700          */
2701         ldlm_pool_del(&ldlm_lock_to_ns(lock)->ns_pool, lock);
2702
2703         /* Consider downgraded lock as a new lock and clear all states
2704          * related to a previous blocking AST processing.
2705          */
2706         ldlm_clear_blocking_data(lock);
2707
2708         lock->l_req_mode = new_mode;
2709         ldlm_grant_lock(lock, NULL);
2710         unlock_res_and_lock(lock);
2711
2712         ldlm_reprocess_all(lock->l_resource, lock);
2713
2714         EXIT;
2715 #endif
2716 }
2717 EXPORT_SYMBOL(ldlm_lock_mode_downgrade);
2718
2719 /**
2720  * Print lock with lock handle \a lockh description into debug log.
2721  *
2722  * Used when printing all locks on a resource for debug purposes.
2723  */
2724 void ldlm_lock_dump_handle(int level, const struct lustre_handle *lockh)
2725 {
2726         struct ldlm_lock *lock;
2727
2728         if (!((libcfs_debug | D_ERROR) & level))
2729                 return;
2730
2731         lock = ldlm_handle2lock(lockh);
2732         if (lock == NULL)
2733                 return;
2734
2735         LDLM_DEBUG_LIMIT(level, lock, "###");
2736
2737         LDLM_LOCK_PUT(lock);
2738 }
2739 EXPORT_SYMBOL(ldlm_lock_dump_handle);
2740
2741 /**
2742  * Print lock information with custom message into debug log.
2743  * Helper function.
2744  */
2745 void _ldlm_lock_debug(struct ldlm_lock *lock,
2746                       struct libcfs_debug_msg_data *msgdata,
2747                       const char *fmt, ...)
2748 {
2749         va_list args;
2750         struct obd_export *exp = lock->l_export;
2751         struct ldlm_resource *resource = NULL;
2752         struct va_format vaf;
2753         char *nid = "local";
2754
2755         /* on server-side resource of lock doesn't change */
2756         if ((lock->l_flags & LDLM_FL_NS_SRV) != 0) {
2757                 if (lock->l_resource != NULL)
2758                         resource = ldlm_resource_getref(lock->l_resource);
2759         } else if (spin_trylock(&lock->l_lock)) {
2760                 if (lock->l_resource != NULL)
2761                         resource = ldlm_resource_getref(lock->l_resource);
2762                 spin_unlock(&lock->l_lock);
2763         }
2764
2765         va_start(args, fmt);
2766         vaf.fmt = fmt;
2767         vaf.va = &args;
2768
2769         if (exp && exp->exp_connection) {
2770                 nid = obd_export_nid2str(exp);
2771         } else if (exp && exp->exp_obd != NULL) {
2772                 struct obd_import *imp = exp->exp_obd->u.cli.cl_import;
2773                 nid = obd_import_nid2str(imp);
2774         }
2775
2776         if (resource == NULL) {
2777                 libcfs_debug_msg(msgdata,
2778                                  "%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",
2779                                  &vaf,
2780                                  lock,
2781                                  lock->l_handle.h_cookie,
2782                                  refcount_read(&lock->l_handle.h_ref),
2783                                  lock->l_readers, lock->l_writers,
2784                                  ldlm_lockname[lock->l_granted_mode],
2785                                  ldlm_lockname[lock->l_req_mode],
2786                                  lock->l_flags, nid,
2787                                  lock->l_remote_handle.cookie,
2788                                  exp ? refcount_read(&exp->exp_handle.h_ref) : -99,
2789                                  lock->l_pid, lock->l_callback_timeout,
2790                                  lock->l_lvb_type);
2791                 va_end(args);
2792                 return;
2793         }
2794
2795         switch (resource->lr_type) {
2796         case LDLM_EXTENT:
2797                 libcfs_debug_msg(msgdata,
2798                                  "%pV ns: %s lock: %p/%#llx lrc: %d/%d,%d mode: %s/%s res: " DLDLMRES " rrc: %d type: %s [%llu->%llu] (req %llu->%llu) flags: %#llx nid: %s remote: %#llx expref: %d pid: %u timeout: %lld lvb_type: %d\n",
2799                                  &vaf,
2800                                  ldlm_lock_to_ns_name(lock), lock,
2801                                  lock->l_handle.h_cookie,
2802                                  refcount_read(&lock->l_handle.h_ref),
2803                                  lock->l_readers, lock->l_writers,
2804                                  ldlm_lockname[lock->l_granted_mode],
2805                                  ldlm_lockname[lock->l_req_mode],
2806                                  PLDLMRES(resource),
2807                                  atomic_read(&resource->lr_refcount),
2808                                  ldlm_typename[resource->lr_type],
2809                                  lock->l_policy_data.l_extent.start,
2810                                  lock->l_policy_data.l_extent.end,
2811                                  lock->l_req_extent.start, lock->l_req_extent.end,
2812                                  lock->l_flags, nid,
2813                                  lock->l_remote_handle.cookie,
2814                                  exp ? refcount_read(&exp->exp_handle.h_ref) : -99,
2815                                  lock->l_pid, lock->l_callback_timeout,
2816                                  lock->l_lvb_type);
2817                 break;
2818
2819         case LDLM_FLOCK:
2820                 libcfs_debug_msg(msgdata,
2821                                  "%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",
2822                                  &vaf,
2823                                  ldlm_lock_to_ns_name(lock), lock,
2824                                  lock->l_handle.h_cookie,
2825                                  refcount_read(&lock->l_handle.h_ref),
2826                                  lock->l_readers, lock->l_writers,
2827                                  ldlm_lockname[lock->l_granted_mode],
2828                                  ldlm_lockname[lock->l_req_mode],
2829                                  PLDLMRES(resource),
2830                                  atomic_read(&resource->lr_refcount),
2831                                  ldlm_typename[resource->lr_type],
2832                                  lock->l_policy_data.l_flock.pid,
2833                                  lock->l_policy_data.l_flock.start,
2834                                  lock->l_policy_data.l_flock.end,
2835                                  lock->l_flags, nid,
2836                                  lock->l_remote_handle.cookie,
2837                                  exp ? refcount_read(&exp->exp_handle.h_ref) : -99,
2838                                  lock->l_pid, lock->l_callback_timeout);
2839                 break;
2840
2841         case LDLM_IBITS:
2842                 libcfs_debug_msg(msgdata,
2843                                  "%pV ns: %s lock: %p/%#llx lrc: %d/%d,%d mode: %s/%s res: " DLDLMRES " bits %#llx/%#llx rrc: %d type: %s flags: %#llx nid: %s remote: %#llx expref: %d pid: %u timeout: %lld lvb_type: %d\n",
2844                                  &vaf,
2845                                  ldlm_lock_to_ns_name(lock),
2846                                  lock, lock->l_handle.h_cookie,
2847                                  refcount_read(&lock->l_handle.h_ref),
2848                                  lock->l_readers, lock->l_writers,
2849                                  ldlm_lockname[lock->l_granted_mode],
2850                                  ldlm_lockname[lock->l_req_mode],
2851                                  PLDLMRES(resource),
2852                                  lock->l_policy_data.l_inodebits.bits,
2853                                  lock->l_policy_data.l_inodebits.try_bits,
2854                                  atomic_read(&resource->lr_refcount),
2855                                  ldlm_typename[resource->lr_type],
2856                                  lock->l_flags, nid,
2857                                  lock->l_remote_handle.cookie,
2858                                  exp ? refcount_read(&exp->exp_handle.h_ref) : -99,
2859                                  lock->l_pid, lock->l_callback_timeout,
2860                                  lock->l_lvb_type);
2861                 break;
2862
2863         default:
2864                 libcfs_debug_msg(msgdata,
2865                                  "%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",
2866                                  &vaf,
2867                                  ldlm_lock_to_ns_name(lock),
2868                                  lock, lock->l_handle.h_cookie,
2869                                  refcount_read(&lock->l_handle.h_ref),
2870                                  lock->l_readers, lock->l_writers,
2871                                  ldlm_lockname[lock->l_granted_mode],
2872                                  ldlm_lockname[lock->l_req_mode],
2873                                  PLDLMRES(resource),
2874                                  atomic_read(&resource->lr_refcount),
2875                                  ldlm_typename[resource->lr_type],
2876                                  lock->l_flags, nid,
2877                                  lock->l_remote_handle.cookie,
2878                                  exp ? refcount_read(&exp->exp_handle.h_ref) : -99,
2879                                  lock->l_pid, lock->l_callback_timeout,
2880                                  lock->l_lvb_type);
2881                 break;
2882         }
2883         va_end(args);
2884         ldlm_resource_putref(resource);
2885 }
2886 EXPORT_SYMBOL(_ldlm_lock_debug);