Whamcloud - gitweb
ORNL-7 replace statahead hacking with its own cache
[fs/lustre-release.git] / lustre / ldlm / ldlm_lock.c
1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2  * vim:expandtab:shiftwidth=8:tabstop=8:
3  *
4  * GPL HEADER START
5  *
6  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
7  *
8  * This program is free software; you can redistribute it and/or modify
9  * it under the terms of the GNU General Public License version 2 only,
10  * as published by the Free Software Foundation.
11  *
12  * This program is distributed in the hope that it will be useful, but
13  * WITHOUT ANY WARRANTY; without even the implied warranty of
14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
15  * General Public License version 2 for more details (a copy is included
16  * in the LICENSE file that accompanied this code).
17  *
18  * You should have received a copy of the GNU General Public License
19  * version 2 along with this program; If not, see
20  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
21  *
22  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
23  * CA 95054 USA or visit www.sun.com if you need additional information or
24  * have any questions.
25  *
26  * GPL HEADER END
27  */
28 /*
29  * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
30  * Use is subject to license terms.
31  *
32  * Copyright (c) 2011 Whamcloud, Inc.
33  *
34  */
35 /*
36  * This file is part of Lustre, http://www.lustre.org/
37  * Lustre is a trademark of Sun Microsystems, Inc.
38  *
39  * lustre/ldlm/ldlm_lock.c
40  *
41  * Author: Peter Braam <braam@clusterfs.com>
42  * Author: Phil Schwan <phil@clusterfs.com>
43  */
44
45 #define DEBUG_SUBSYSTEM S_LDLM
46
47 #ifdef __KERNEL__
48 # include <libcfs/libcfs.h>
49 # include <linux/lustre_intent.h>
50 #else
51 # include <liblustre.h>
52 #endif
53
54 #include <obd_class.h>
55 #include "ldlm_internal.h"
56
57 /* lock types */
58 char *ldlm_lockname[] = {
59         [0] "--",
60         [LCK_EX] "EX",
61         [LCK_PW] "PW",
62         [LCK_PR] "PR",
63         [LCK_CW] "CW",
64         [LCK_CR] "CR",
65         [LCK_NL] "NL",
66         [LCK_GROUP] "GROUP",
67         [LCK_COS] "COS"
68 };
69
70 char *ldlm_typename[] = {
71         [LDLM_PLAIN] "PLN",
72         [LDLM_EXTENT] "EXT",
73         [LDLM_FLOCK] "FLK",
74         [LDLM_IBITS] "IBT",
75 };
76
77 static ldlm_policy_wire_to_local_t ldlm_policy_wire_to_local[] = {
78         [LDLM_PLAIN - LDLM_MIN_TYPE] ldlm_plain_policy_wire_to_local,
79         [LDLM_EXTENT - LDLM_MIN_TYPE] ldlm_extent_policy_wire_to_local,
80         [LDLM_FLOCK - LDLM_MIN_TYPE] ldlm_flock_policy_wire_to_local,
81         [LDLM_IBITS - LDLM_MIN_TYPE] ldlm_ibits_policy_wire_to_local,
82 };
83
84 static ldlm_policy_local_to_wire_t ldlm_policy_local_to_wire[] = {
85         [LDLM_PLAIN - LDLM_MIN_TYPE] ldlm_plain_policy_local_to_wire,
86         [LDLM_EXTENT - LDLM_MIN_TYPE] ldlm_extent_policy_local_to_wire,
87         [LDLM_FLOCK - LDLM_MIN_TYPE] ldlm_flock_policy_local_to_wire,
88         [LDLM_IBITS - LDLM_MIN_TYPE] ldlm_ibits_policy_local_to_wire,
89 };
90
91 /**
92  * Converts lock policy from local format to on the wire lock_desc format
93  */
94 void ldlm_convert_policy_to_wire(ldlm_type_t type,
95                                  const ldlm_policy_data_t *lpolicy,
96                                  ldlm_wire_policy_data_t *wpolicy)
97 {
98         ldlm_policy_local_to_wire_t convert;
99
100         convert = ldlm_policy_local_to_wire[type - LDLM_MIN_TYPE];
101
102         convert(lpolicy, wpolicy);
103 }
104
105 /**
106  * Converts lock policy from on the wire lock_desc format to local format
107  */
108 void ldlm_convert_policy_to_local(ldlm_type_t type,
109                                   const ldlm_wire_policy_data_t *wpolicy,
110                                   ldlm_policy_data_t *lpolicy)
111 {
112         ldlm_policy_wire_to_local_t convert;
113
114         convert = ldlm_policy_wire_to_local[type - LDLM_MIN_TYPE];
115
116         convert(wpolicy, lpolicy);
117 }
118
119 char *ldlm_it2str(int it)
120 {
121         switch (it) {
122         case IT_OPEN:
123                 return "open";
124         case IT_CREAT:
125                 return "creat";
126         case (IT_OPEN | IT_CREAT):
127                 return "open|creat";
128         case IT_READDIR:
129                 return "readdir";
130         case IT_GETATTR:
131                 return "getattr";
132         case IT_LOOKUP:
133                 return "lookup";
134         case IT_UNLINK:
135                 return "unlink";
136         case IT_GETXATTR:
137                 return "getxattr";
138         default:
139                 CERROR("Unknown intent %d\n", it);
140                 return "UNKNOWN";
141         }
142 }
143
144 extern cfs_mem_cache_t *ldlm_lock_slab;
145
146 static ldlm_processing_policy ldlm_processing_policy_table[] = {
147         [LDLM_PLAIN] ldlm_process_plain_lock,
148         [LDLM_EXTENT] ldlm_process_extent_lock,
149 #ifdef __KERNEL__
150         [LDLM_FLOCK] ldlm_process_flock_lock,
151 #endif
152         [LDLM_IBITS] ldlm_process_inodebits_lock,
153 };
154
155 ldlm_processing_policy ldlm_get_processing_policy(struct ldlm_resource *res)
156 {
157         return ldlm_processing_policy_table[res->lr_type];
158 }
159
160 void ldlm_register_intent(struct ldlm_namespace *ns, ldlm_res_policy arg)
161 {
162         ns->ns_policy = arg;
163 }
164
165 /*
166  * REFCOUNTED LOCK OBJECTS
167  */
168
169
170 /*
171  * Lock refcounts, during creation:
172  *   - one special one for allocation, dec'd only once in destroy
173  *   - one for being a lock that's in-use
174  *   - one for the addref associated with a new lock
175  */
176 struct ldlm_lock *ldlm_lock_get(struct ldlm_lock *lock)
177 {
178         cfs_atomic_inc(&lock->l_refc);
179         return lock;
180 }
181
182 static void ldlm_lock_free(struct ldlm_lock *lock, size_t size)
183 {
184         LASSERT(size == sizeof(*lock));
185         OBD_SLAB_FREE(lock, ldlm_lock_slab, sizeof(*lock));
186 }
187
188 void ldlm_lock_put(struct ldlm_lock *lock)
189 {
190         ENTRY;
191
192         LASSERT(lock->l_resource != LP_POISON);
193         LASSERT(cfs_atomic_read(&lock->l_refc) > 0);
194         if (cfs_atomic_dec_and_test(&lock->l_refc)) {
195                 struct ldlm_resource *res;
196
197                 LDLM_DEBUG(lock,
198                            "final lock_put on destroyed lock, freeing it.");
199
200                 res = lock->l_resource;
201                 LASSERT(lock->l_destroyed);
202                 LASSERT(cfs_list_empty(&lock->l_res_link));
203                 LASSERT(cfs_list_empty(&lock->l_pending_chain));
204
205                 lprocfs_counter_decr(ldlm_res_to_ns(res)->ns_stats,
206                                      LDLM_NSS_LOCKS);
207                 lu_ref_del(&res->lr_reference, "lock", lock);
208                 ldlm_resource_putref(res);
209                 lock->l_resource = NULL;
210                 if (lock->l_export) {
211                         class_export_lock_put(lock->l_export, lock);
212                         lock->l_export = NULL;
213                 }
214
215                 if (lock->l_lvb_data != NULL)
216                         OBD_FREE(lock->l_lvb_data, lock->l_lvb_len);
217
218                 ldlm_interval_free(ldlm_interval_detach(lock));
219                 lu_ref_fini(&lock->l_reference);
220                 OBD_FREE_RCU_CB(lock, sizeof(*lock), &lock->l_handle,
221                                 ldlm_lock_free);
222         }
223
224         EXIT;
225 }
226
227 int ldlm_lock_remove_from_lru_nolock(struct ldlm_lock *lock)
228 {
229         int rc = 0;
230         if (!cfs_list_empty(&lock->l_lru)) {
231                 struct ldlm_namespace *ns = ldlm_lock_to_ns(lock);
232
233                 LASSERT(lock->l_resource->lr_type != LDLM_FLOCK);
234                 cfs_list_del_init(&lock->l_lru);
235                 if (lock->l_flags & LDLM_FL_SKIPPED)
236                         lock->l_flags &= ~LDLM_FL_SKIPPED;
237                 LASSERT(ns->ns_nr_unused > 0);
238                 ns->ns_nr_unused--;
239                 rc = 1;
240         }
241         return rc;
242 }
243
244 int ldlm_lock_remove_from_lru(struct ldlm_lock *lock)
245 {
246         struct ldlm_namespace *ns = ldlm_lock_to_ns(lock);
247         int rc;
248
249         ENTRY;
250         if (lock->l_ns_srv) {
251                 LASSERT(cfs_list_empty(&lock->l_lru));
252                 RETURN(0);
253         }
254
255         cfs_spin_lock(&ns->ns_lock);
256         rc = ldlm_lock_remove_from_lru_nolock(lock);
257         cfs_spin_unlock(&ns->ns_lock);
258         EXIT;
259         return rc;
260 }
261
262 void ldlm_lock_add_to_lru_nolock(struct ldlm_lock *lock)
263 {
264         struct ldlm_namespace *ns = ldlm_lock_to_ns(lock);
265
266         lock->l_last_used = cfs_time_current();
267         LASSERT(cfs_list_empty(&lock->l_lru));
268         LASSERT(lock->l_resource->lr_type != LDLM_FLOCK);
269         cfs_list_add_tail(&lock->l_lru, &ns->ns_unused_list);
270         LASSERT(ns->ns_nr_unused >= 0);
271         ns->ns_nr_unused++;
272 }
273
274 void ldlm_lock_add_to_lru(struct ldlm_lock *lock)
275 {
276         struct ldlm_namespace *ns = ldlm_lock_to_ns(lock);
277
278         ENTRY;
279         cfs_spin_lock(&ns->ns_lock);
280         ldlm_lock_add_to_lru_nolock(lock);
281         cfs_spin_unlock(&ns->ns_lock);
282         EXIT;
283 }
284
285 void ldlm_lock_touch_in_lru(struct ldlm_lock *lock)
286 {
287         struct ldlm_namespace *ns = ldlm_lock_to_ns(lock);
288
289         ENTRY;
290         if (lock->l_ns_srv) {
291                 LASSERT(cfs_list_empty(&lock->l_lru));
292                 EXIT;
293                 return;
294         }
295
296         cfs_spin_lock(&ns->ns_lock);
297         if (!cfs_list_empty(&lock->l_lru)) {
298                 ldlm_lock_remove_from_lru_nolock(lock);
299                 ldlm_lock_add_to_lru_nolock(lock);
300         }
301         cfs_spin_unlock(&ns->ns_lock);
302         EXIT;
303 }
304
305 /* This used to have a 'strict' flag, which recovery would use to mark an
306  * in-use lock as needing-to-die.  Lest I am ever tempted to put it back, I
307  * shall explain why it's gone: with the new hash table scheme, once you call
308  * ldlm_lock_destroy, you can never drop your final references on this lock.
309  * Because it's not in the hash table anymore.  -phil */
310 int ldlm_lock_destroy_internal(struct ldlm_lock *lock)
311 {
312         ENTRY;
313
314         if (lock->l_readers || lock->l_writers) {
315                 LDLM_ERROR(lock, "lock still has references");
316                 ldlm_lock_dump(D_ERROR, lock, 0);
317                 LBUG();
318         }
319
320         if (!cfs_list_empty(&lock->l_res_link)) {
321                 LDLM_ERROR(lock, "lock still on resource");
322                 ldlm_lock_dump(D_ERROR, lock, 0);
323                 LBUG();
324         }
325
326         if (lock->l_destroyed) {
327                 LASSERT(cfs_list_empty(&lock->l_lru));
328                 EXIT;
329                 return 0;
330         }
331         lock->l_destroyed = 1;
332
333         if (lock->l_export && lock->l_export->exp_lock_hash &&
334             !cfs_hlist_unhashed(&lock->l_exp_hash))
335                 cfs_hash_del(lock->l_export->exp_lock_hash,
336                              &lock->l_remote_handle, &lock->l_exp_hash);
337
338         ldlm_lock_remove_from_lru(lock);
339         class_handle_unhash(&lock->l_handle);
340
341 #if 0
342         /* Wake anyone waiting for this lock */
343         /* FIXME: I should probably add yet another flag, instead of using
344          * l_export to only call this on clients */
345         if (lock->l_export)
346                 class_export_put(lock->l_export);
347         lock->l_export = NULL;
348         if (lock->l_export && lock->l_completion_ast)
349                 lock->l_completion_ast(lock, 0);
350 #endif
351         EXIT;
352         return 1;
353 }
354
355 void ldlm_lock_destroy(struct ldlm_lock *lock)
356 {
357         int first;
358         ENTRY;
359         lock_res_and_lock(lock);
360         first = ldlm_lock_destroy_internal(lock);
361         unlock_res_and_lock(lock);
362
363         /* drop reference from hashtable only for first destroy */
364         if (first) {
365                 lu_ref_del(&lock->l_reference, "hash", lock);
366                 LDLM_LOCK_RELEASE(lock);
367         }
368         EXIT;
369 }
370
371 void ldlm_lock_destroy_nolock(struct ldlm_lock *lock)
372 {
373         int first;
374         ENTRY;
375         first = ldlm_lock_destroy_internal(lock);
376         /* drop reference from hashtable only for first destroy */
377         if (first) {
378                 lu_ref_del(&lock->l_reference, "hash", lock);
379                 LDLM_LOCK_RELEASE(lock);
380         }
381         EXIT;
382 }
383
384 /* this is called by portals_handle2object with the handle lock taken */
385 static void lock_handle_addref(void *lock)
386 {
387         LDLM_LOCK_GET((struct ldlm_lock *)lock);
388 }
389
390 /*
391  * usage: pass in a resource on which you have done ldlm_resource_get
392  *        new lock will take over the refcount.
393  * returns: lock with refcount 2 - one for current caller and one for remote
394  */
395 static struct ldlm_lock *ldlm_lock_new(struct ldlm_resource *resource)
396 {
397         struct ldlm_lock *lock;
398         ENTRY;
399
400         if (resource == NULL)
401                 LBUG();
402
403         OBD_SLAB_ALLOC_PTR_GFP(lock, ldlm_lock_slab, CFS_ALLOC_IO);
404         if (lock == NULL)
405                 RETURN(NULL);
406
407         cfs_spin_lock_init(&lock->l_lock);
408         lock->l_resource = resource;
409         lu_ref_add(&resource->lr_reference, "lock", lock);
410
411         cfs_atomic_set(&lock->l_refc, 2);
412         CFS_INIT_LIST_HEAD(&lock->l_res_link);
413         CFS_INIT_LIST_HEAD(&lock->l_lru);
414         CFS_INIT_LIST_HEAD(&lock->l_pending_chain);
415         CFS_INIT_LIST_HEAD(&lock->l_bl_ast);
416         CFS_INIT_LIST_HEAD(&lock->l_cp_ast);
417         CFS_INIT_LIST_HEAD(&lock->l_rk_ast);
418         cfs_waitq_init(&lock->l_waitq);
419         lock->l_blocking_lock = NULL;
420         CFS_INIT_LIST_HEAD(&lock->l_sl_mode);
421         CFS_INIT_LIST_HEAD(&lock->l_sl_policy);
422         CFS_INIT_HLIST_NODE(&lock->l_exp_hash);
423
424         lprocfs_counter_incr(ldlm_res_to_ns(resource)->ns_stats,
425                              LDLM_NSS_LOCKS);
426         CFS_INIT_LIST_HEAD(&lock->l_handle.h_link);
427         class_handle_hash(&lock->l_handle, lock_handle_addref);
428
429         lu_ref_init(&lock->l_reference);
430         lu_ref_add(&lock->l_reference, "hash", lock);
431         lock->l_callback_timeout = 0;
432
433 #if LUSTRE_TRACKS_LOCK_EXP_REFS
434         CFS_INIT_LIST_HEAD(&lock->l_exp_refs_link);
435         lock->l_exp_refs_nr = 0;
436         lock->l_exp_refs_target = NULL;
437 #endif
438
439         RETURN(lock);
440 }
441
442 int ldlm_lock_change_resource(struct ldlm_namespace *ns, struct ldlm_lock *lock,
443                               const struct ldlm_res_id *new_resid)
444 {
445         struct ldlm_resource *oldres = lock->l_resource;
446         struct ldlm_resource *newres;
447         int type;
448         ENTRY;
449
450         LASSERT(ns_is_client(ns));
451
452         lock_res_and_lock(lock);
453         if (memcmp(new_resid, &lock->l_resource->lr_name,
454                    sizeof(lock->l_resource->lr_name)) == 0) {
455                 /* Nothing to do */
456                 unlock_res_and_lock(lock);
457                 RETURN(0);
458         }
459
460         LASSERT(new_resid->name[0] != 0);
461
462         /* This function assumes that the lock isn't on any lists */
463         LASSERT(cfs_list_empty(&lock->l_res_link));
464
465         type = oldres->lr_type;
466         unlock_res_and_lock(lock);
467
468         newres = ldlm_resource_get(ns, NULL, new_resid, type, 1);
469         if (newres == NULL)
470                 RETURN(-ENOMEM);
471
472         lu_ref_add(&newres->lr_reference, "lock", lock);
473         /*
474          * To flip the lock from the old to the new resource, lock, oldres and
475          * newres have to be locked. Resource spin-locks are nested within
476          * lock->l_lock, and are taken in the memory address order to avoid
477          * dead-locks.
478          */
479         cfs_spin_lock(&lock->l_lock);
480         oldres = lock->l_resource;
481         if (oldres < newres) {
482                 lock_res(oldres);
483                 lock_res_nested(newres, LRT_NEW);
484         } else {
485                 lock_res(newres);
486                 lock_res_nested(oldres, LRT_NEW);
487         }
488         LASSERT(memcmp(new_resid, &oldres->lr_name,
489                        sizeof oldres->lr_name) != 0);
490         lock->l_resource = newres;
491         unlock_res(oldres);
492         unlock_res_and_lock(lock);
493
494         /* ...and the flowers are still standing! */
495         lu_ref_del(&oldres->lr_reference, "lock", lock);
496         ldlm_resource_putref(oldres);
497
498         RETURN(0);
499 }
500
501 /*
502  *  HANDLES
503  */
504
505 void ldlm_lock2handle(const struct ldlm_lock *lock, struct lustre_handle *lockh)
506 {
507         lockh->cookie = lock->l_handle.h_cookie;
508 }
509
510 /* if flags: atomically get the lock and set the flags.
511  *           Return NULL if flag already set
512  */
513
514 struct ldlm_lock *__ldlm_handle2lock(const struct lustre_handle *handle,
515                                      int flags)
516 {
517         struct ldlm_lock *lock;
518         ENTRY;
519
520         LASSERT(handle);
521
522         lock = class_handle2object(handle->cookie);
523         if (lock == NULL)
524                 RETURN(NULL);
525
526         /* It's unlikely but possible that someone marked the lock as
527          * destroyed after we did handle2object on it */
528         if (flags == 0 && !lock->l_destroyed) {
529                 lu_ref_add(&lock->l_reference, "handle", cfs_current());
530                 RETURN(lock);
531         }
532
533         lock_res_and_lock(lock);
534
535         LASSERT(lock->l_resource != NULL);
536
537         lu_ref_add_atomic(&lock->l_reference, "handle", cfs_current());
538         if (unlikely(lock->l_destroyed)) {
539                 unlock_res_and_lock(lock);
540                 CDEBUG(D_INFO, "lock already destroyed: lock %p\n", lock);
541                 LDLM_LOCK_PUT(lock);
542                 RETURN(NULL);
543         }
544
545         if (flags && (lock->l_flags & flags)) {
546                 unlock_res_and_lock(lock);
547                 LDLM_LOCK_PUT(lock);
548                 RETURN(NULL);
549         }
550
551         if (flags)
552                 lock->l_flags |= flags;
553
554         unlock_res_and_lock(lock);
555         RETURN(lock);
556 }
557
558 void ldlm_lock2desc(struct ldlm_lock *lock, struct ldlm_lock_desc *desc)
559 {
560         struct obd_export *exp = lock->l_export?:lock->l_conn_export;
561         /* INODEBITS_INTEROP: If the other side does not support
562          * inodebits, reply with a plain lock descriptor.
563          */
564         if ((lock->l_resource->lr_type == LDLM_IBITS) &&
565             (exp && !(exp->exp_connect_flags & OBD_CONNECT_IBITS))) {
566                 /* Make sure all the right bits are set in this lock we
567                    are going to pass to client */
568                 LASSERTF(lock->l_policy_data.l_inodebits.bits ==
569                          (MDS_INODELOCK_LOOKUP|MDS_INODELOCK_UPDATE),
570                          "Inappropriate inode lock bits during "
571                          "conversion " LPU64 "\n",
572                          lock->l_policy_data.l_inodebits.bits);
573
574                 ldlm_res2desc(lock->l_resource, &desc->l_resource);
575                 desc->l_resource.lr_type = LDLM_PLAIN;
576
577                 /* Convert "new" lock mode to something old client can
578                    understand */
579                 if ((lock->l_req_mode == LCK_CR) ||
580                     (lock->l_req_mode == LCK_CW))
581                         desc->l_req_mode = LCK_PR;
582                 else
583                         desc->l_req_mode = lock->l_req_mode;
584                 if ((lock->l_granted_mode == LCK_CR) ||
585                     (lock->l_granted_mode == LCK_CW)) {
586                         desc->l_granted_mode = LCK_PR;
587                 } else {
588                         /* We never grant PW/EX locks to clients */
589                         LASSERT((lock->l_granted_mode != LCK_PW) &&
590                                 (lock->l_granted_mode != LCK_EX));
591                         desc->l_granted_mode = lock->l_granted_mode;
592                 }
593
594                 /* We do not copy policy here, because there is no
595                    policy for plain locks */
596         } else {
597                 ldlm_res2desc(lock->l_resource, &desc->l_resource);
598                 desc->l_req_mode = lock->l_req_mode;
599                 desc->l_granted_mode = lock->l_granted_mode;
600                 ldlm_convert_policy_to_wire(lock->l_resource->lr_type,
601                                             &lock->l_policy_data,
602                                             &desc->l_policy_data);
603         }
604 }
605
606 void ldlm_add_bl_work_item(struct ldlm_lock *lock, struct ldlm_lock *new,
607                            cfs_list_t *work_list)
608 {
609         if ((lock->l_flags & LDLM_FL_AST_SENT) == 0) {
610                 LDLM_DEBUG(lock, "lock incompatible; sending blocking AST.");
611                 lock->l_flags |= LDLM_FL_AST_SENT;
612                 /* If the enqueuing client said so, tell the AST recipient to
613                  * discard dirty data, rather than writing back. */
614                 if (new->l_flags & LDLM_AST_DISCARD_DATA)
615                         lock->l_flags |= LDLM_FL_DISCARD_DATA;
616                 LASSERT(cfs_list_empty(&lock->l_bl_ast));
617                 cfs_list_add(&lock->l_bl_ast, work_list);
618                 LDLM_LOCK_GET(lock);
619                 LASSERT(lock->l_blocking_lock == NULL);
620                 lock->l_blocking_lock = LDLM_LOCK_GET(new);
621         }
622 }
623
624 void ldlm_add_cp_work_item(struct ldlm_lock *lock, cfs_list_t *work_list)
625 {
626         if ((lock->l_flags & LDLM_FL_CP_REQD) == 0) {
627                 lock->l_flags |= LDLM_FL_CP_REQD;
628                 LDLM_DEBUG(lock, "lock granted; sending completion AST.");
629                 LASSERT(cfs_list_empty(&lock->l_cp_ast));
630                 cfs_list_add(&lock->l_cp_ast, work_list);
631                 LDLM_LOCK_GET(lock);
632         }
633 }
634
635 /* must be called with lr_lock held */
636 void ldlm_add_ast_work_item(struct ldlm_lock *lock, struct ldlm_lock *new,
637                             cfs_list_t *work_list)
638 {
639         ENTRY;
640         check_res_locked(lock->l_resource);
641         if (new)
642                 ldlm_add_bl_work_item(lock, new, work_list);
643         else
644                 ldlm_add_cp_work_item(lock, work_list);
645         EXIT;
646 }
647
648 void ldlm_lock_addref(struct lustre_handle *lockh, __u32 mode)
649 {
650         struct ldlm_lock *lock;
651
652         lock = ldlm_handle2lock(lockh);
653         LASSERT(lock != NULL);
654         ldlm_lock_addref_internal(lock, mode);
655         LDLM_LOCK_PUT(lock);
656 }
657
658 void ldlm_lock_addref_internal_nolock(struct ldlm_lock *lock, __u32 mode)
659 {
660         ldlm_lock_remove_from_lru(lock);
661         if (mode & (LCK_NL | LCK_CR | LCK_PR)) {
662                 lock->l_readers++;
663                 lu_ref_add_atomic(&lock->l_reference, "reader", lock);
664         }
665         if (mode & (LCK_EX | LCK_CW | LCK_PW | LCK_GROUP | LCK_COS)) {
666                 lock->l_writers++;
667                 lu_ref_add_atomic(&lock->l_reference, "writer", lock);
668         }
669         LDLM_LOCK_GET(lock);
670         lu_ref_add_atomic(&lock->l_reference, "user", lock);
671         LDLM_DEBUG(lock, "ldlm_lock_addref(%s)", ldlm_lockname[mode]);
672 }
673
674 /**
675  * Attempts to addref a lock, and fails if lock is already LDLM_FL_CBPENDING
676  * or destroyed.
677  *
678  * \retval 0 success, lock was addref-ed
679  *
680  * \retval -EAGAIN lock is being canceled.
681  */
682 int ldlm_lock_addref_try(struct lustre_handle *lockh, __u32 mode)
683 {
684         struct ldlm_lock *lock;
685         int               result;
686
687         result = -EAGAIN;
688         lock = ldlm_handle2lock(lockh);
689         if (lock != NULL) {
690                 lock_res_and_lock(lock);
691                 if (lock->l_readers != 0 || lock->l_writers != 0 ||
692                     !(lock->l_flags & LDLM_FL_CBPENDING)) {
693                         ldlm_lock_addref_internal_nolock(lock, mode);
694                         result = 0;
695                 }
696                 unlock_res_and_lock(lock);
697                 LDLM_LOCK_PUT(lock);
698         }
699         return result;
700 }
701
702 /* only called for local locks */
703 void ldlm_lock_addref_internal(struct ldlm_lock *lock, __u32 mode)
704 {
705         lock_res_and_lock(lock);
706         ldlm_lock_addref_internal_nolock(lock, mode);
707         unlock_res_and_lock(lock);
708 }
709
710 /* only called in ldlm_flock_destroy and for local locks.
711  *  * for LDLM_FLOCK type locks, l_blocking_ast is null, and
712  *   * ldlm_lock_remove_from_lru() does nothing, it is safe
713  *    * for ldlm_flock_destroy usage by dropping some code */
714 void ldlm_lock_decref_internal_nolock(struct ldlm_lock *lock, __u32 mode)
715 {
716         LDLM_DEBUG(lock, "ldlm_lock_decref(%s)", ldlm_lockname[mode]);
717         if (mode & (LCK_NL | LCK_CR | LCK_PR)) {
718                 LASSERT(lock->l_readers > 0);
719                 lu_ref_del(&lock->l_reference, "reader", lock);
720                 lock->l_readers--;
721         }
722         if (mode & (LCK_EX | LCK_CW | LCK_PW | LCK_GROUP | LCK_COS)) {
723                 LASSERT(lock->l_writers > 0);
724                 lu_ref_del(&lock->l_reference, "writer", lock);
725                 lock->l_writers--;
726         }
727
728         lu_ref_del(&lock->l_reference, "user", lock);
729         LDLM_LOCK_RELEASE(lock);    /* matches the LDLM_LOCK_GET() in addref */
730 }
731
732 void ldlm_lock_decref_internal(struct ldlm_lock *lock, __u32 mode)
733 {
734         struct ldlm_namespace *ns;
735         ENTRY;
736
737         lock_res_and_lock(lock);
738
739         ns = ldlm_lock_to_ns(lock);
740
741         ldlm_lock_decref_internal_nolock(lock, mode);
742
743         if (lock->l_flags & LDLM_FL_LOCAL &&
744             !lock->l_readers && !lock->l_writers) {
745                 /* If this is a local lock on a server namespace and this was
746                  * the last reference, cancel the lock. */
747                 CDEBUG(D_INFO, "forcing cancel of local lock\n");
748                 lock->l_flags |= LDLM_FL_CBPENDING;
749         }
750
751         if (!lock->l_readers && !lock->l_writers &&
752             (lock->l_flags & LDLM_FL_CBPENDING)) {
753                 /* If we received a blocked AST and this was the last reference,
754                  * run the callback. */
755                 if (lock->l_ns_srv && lock->l_export)
756                         CERROR("FL_CBPENDING set on non-local lock--just a "
757                                "warning\n");
758
759                 LDLM_DEBUG(lock, "final decref done on cbpending lock");
760
761                 LDLM_LOCK_GET(lock); /* dropped by bl thread */
762                 ldlm_lock_remove_from_lru(lock);
763                 unlock_res_and_lock(lock);
764
765                 if (lock->l_flags & LDLM_FL_FAIL_LOC)
766                         OBD_RACE(OBD_FAIL_LDLM_CP_BL_RACE);
767
768                 if ((lock->l_flags & LDLM_FL_ATOMIC_CB) ||
769                     ldlm_bl_to_thread_lock(ns, NULL, lock) != 0)
770                         ldlm_handle_bl_callback(ns, NULL, lock);
771         } else if (ns_is_client(ns) &&
772                    !lock->l_readers && !lock->l_writers &&
773                    !(lock->l_flags & LDLM_FL_NO_LRU) &&
774                    !(lock->l_flags & LDLM_FL_BL_AST)) {
775
776                 LDLM_DEBUG(lock, "add lock into lru list");
777
778                 /* If this is a client-side namespace and this was the last
779                  * reference, put it on the LRU. */
780                 ldlm_lock_add_to_lru(lock);
781                 unlock_res_and_lock(lock);
782
783                 if (lock->l_flags & LDLM_FL_FAIL_LOC)
784                         OBD_RACE(OBD_FAIL_LDLM_CP_BL_RACE);
785
786                 /* Call ldlm_cancel_lru() only if EARLY_CANCEL and LRU RESIZE
787                  * are not supported by the server, otherwise, it is done on
788                  * enqueue. */
789                 if (!exp_connect_cancelset(lock->l_conn_export) &&
790                     !ns_connect_lru_resize(ns))
791                         ldlm_cancel_lru(ns, 0, LDLM_ASYNC, 0);
792         } else {
793                 LDLM_DEBUG(lock, "do not add lock into lru list");
794                 unlock_res_and_lock(lock);
795         }
796
797         EXIT;
798 }
799
800 void ldlm_lock_decref(struct lustre_handle *lockh, __u32 mode)
801 {
802         struct ldlm_lock *lock = __ldlm_handle2lock(lockh, 0);
803         LASSERTF(lock != NULL, "Non-existing lock: "LPX64"\n", lockh->cookie);
804         ldlm_lock_decref_internal(lock, mode);
805         LDLM_LOCK_PUT(lock);
806 }
807
808 /* This will drop a lock reference and mark it for destruction, but will not
809  * necessarily cancel the lock before returning. */
810 void ldlm_lock_decref_and_cancel(struct lustre_handle *lockh, __u32 mode)
811 {
812         struct ldlm_lock *lock = __ldlm_handle2lock(lockh, 0);
813         ENTRY;
814
815         LASSERT(lock != NULL);
816
817         LDLM_DEBUG(lock, "ldlm_lock_decref(%s)", ldlm_lockname[mode]);
818         lock_res_and_lock(lock);
819         lock->l_flags |= LDLM_FL_CBPENDING;
820         unlock_res_and_lock(lock);
821         ldlm_lock_decref_internal(lock, mode);
822         LDLM_LOCK_PUT(lock);
823 }
824
825 struct sl_insert_point {
826         cfs_list_t *res_link;
827         cfs_list_t *mode_link;
828         cfs_list_t *policy_link;
829 };
830
831 /*
832  * search_granted_lock
833  *
834  * Description:
835  *      Finds a position to insert the new lock.
836  * Parameters:
837  *      queue [input]:  the granted list where search acts on;
838  *      req [input]:    the lock whose position to be located;
839  *      prev [output]:  positions within 3 lists to insert @req to
840  * Return Value:
841  *      filled @prev
842  * NOTE: called by
843  *  - ldlm_grant_lock_with_skiplist
844  */
845 static void search_granted_lock(cfs_list_t *queue,
846                                 struct ldlm_lock *req,
847                                 struct sl_insert_point *prev)
848 {
849         cfs_list_t *tmp;
850         struct ldlm_lock *lock, *mode_end, *policy_end;
851         ENTRY;
852
853         cfs_list_for_each(tmp, queue) {
854                 lock = cfs_list_entry(tmp, struct ldlm_lock, l_res_link);
855
856                 mode_end = cfs_list_entry(lock->l_sl_mode.prev,
857                                           struct ldlm_lock, l_sl_mode);
858
859                 if (lock->l_req_mode != req->l_req_mode) {
860                         /* jump to last lock of mode group */
861                         tmp = &mode_end->l_res_link;
862                         continue;
863                 }
864
865                 /* suitable mode group is found */
866                 if (lock->l_resource->lr_type == LDLM_PLAIN) {
867                         /* insert point is last lock of the mode group */
868                         prev->res_link = &mode_end->l_res_link;
869                         prev->mode_link = &mode_end->l_sl_mode;
870                         prev->policy_link = &req->l_sl_policy;
871                         EXIT;
872                         return;
873                 } else if (lock->l_resource->lr_type == LDLM_IBITS) {
874                         for (;;) {
875                                 policy_end =
876                                         cfs_list_entry(lock->l_sl_policy.prev,
877                                                        struct ldlm_lock,
878                                                        l_sl_policy);
879
880                                 if (lock->l_policy_data.l_inodebits.bits ==
881                                     req->l_policy_data.l_inodebits.bits) {
882                                         /* insert point is last lock of
883                                          * the policy group */
884                                         prev->res_link =
885                                                 &policy_end->l_res_link;
886                                         prev->mode_link =
887                                                 &policy_end->l_sl_mode;
888                                         prev->policy_link =
889                                                 &policy_end->l_sl_policy;
890                                         EXIT;
891                                         return;
892                                 }
893
894                                 if (policy_end == mode_end)
895                                         /* done with mode group */
896                                         break;
897
898                                 /* go to next policy group within mode group */
899                                 tmp = policy_end->l_res_link.next;
900                                 lock = cfs_list_entry(tmp, struct ldlm_lock,
901                                                       l_res_link);
902                         }  /* loop over policy groups within the mode group */
903
904                         /* insert point is last lock of the mode group,
905                          * new policy group is started */
906                         prev->res_link = &mode_end->l_res_link;
907                         prev->mode_link = &mode_end->l_sl_mode;
908                         prev->policy_link = &req->l_sl_policy;
909                         EXIT;
910                         return;
911                 } else {
912                         LDLM_ERROR(lock,"is not LDLM_PLAIN or LDLM_IBITS lock");
913                         LBUG();
914                 }
915         }
916
917         /* insert point is last lock on the queue,
918          * new mode group and new policy group are started */
919         prev->res_link = queue->prev;
920         prev->mode_link = &req->l_sl_mode;
921         prev->policy_link = &req->l_sl_policy;
922         EXIT;
923         return;
924 }
925
926 static void ldlm_granted_list_add_lock(struct ldlm_lock *lock,
927                                        struct sl_insert_point *prev)
928 {
929         struct ldlm_resource *res = lock->l_resource;
930         ENTRY;
931
932         check_res_locked(res);
933
934         ldlm_resource_dump(D_INFO, res);
935         CDEBUG(D_OTHER, "About to add this lock:\n");
936         ldlm_lock_dump(D_OTHER, lock, 0);
937
938         if (lock->l_destroyed) {
939                 CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n");
940                 return;
941         }
942
943         LASSERT(cfs_list_empty(&lock->l_res_link));
944         LASSERT(cfs_list_empty(&lock->l_sl_mode));
945         LASSERT(cfs_list_empty(&lock->l_sl_policy));
946
947         cfs_list_add(&lock->l_res_link, prev->res_link);
948         cfs_list_add(&lock->l_sl_mode, prev->mode_link);
949         cfs_list_add(&lock->l_sl_policy, prev->policy_link);
950
951         EXIT;
952 }
953
954 static void ldlm_grant_lock_with_skiplist(struct ldlm_lock *lock)
955 {
956         struct sl_insert_point prev;
957         ENTRY;
958
959         LASSERT(lock->l_req_mode == lock->l_granted_mode);
960
961         search_granted_lock(&lock->l_resource->lr_granted, lock, &prev);
962         ldlm_granted_list_add_lock(lock, &prev);
963         EXIT;
964 }
965
966 /* NOTE: called by
967  *  - ldlm_lock_enqueue
968  *  - ldlm_reprocess_queue
969  *  - ldlm_lock_convert
970  *
971  * must be called with lr_lock held
972  */
973 void ldlm_grant_lock(struct ldlm_lock *lock, cfs_list_t *work_list)
974 {
975         struct ldlm_resource *res = lock->l_resource;
976         ENTRY;
977
978         check_res_locked(res);
979
980         lock->l_granted_mode = lock->l_req_mode;
981         if (res->lr_type == LDLM_PLAIN || res->lr_type == LDLM_IBITS)
982                 ldlm_grant_lock_with_skiplist(lock);
983         else if (res->lr_type == LDLM_EXTENT)
984                 ldlm_extent_add_lock(res, lock);
985         else
986                 ldlm_resource_add_lock(res, &res->lr_granted, lock);
987
988         if (lock->l_granted_mode < res->lr_most_restr)
989                 res->lr_most_restr = lock->l_granted_mode;
990
991         if (work_list && lock->l_completion_ast != NULL)
992                 ldlm_add_ast_work_item(lock, NULL, work_list);
993
994         ldlm_pool_add(&ldlm_res_to_ns(res)->ns_pool, lock);
995         EXIT;
996 }
997
998 /* returns a referenced lock or NULL.  See the flag descriptions below, in the
999  * comment above ldlm_lock_match */
1000 static struct ldlm_lock *search_queue(cfs_list_t *queue,
1001                                       ldlm_mode_t *mode,
1002                                       ldlm_policy_data_t *policy,
1003                                       struct ldlm_lock *old_lock,
1004                                       int flags, int unref)
1005 {
1006         struct ldlm_lock *lock;
1007         cfs_list_t       *tmp;
1008
1009         cfs_list_for_each(tmp, queue) {
1010                 ldlm_mode_t match;
1011
1012                 lock = cfs_list_entry(tmp, struct ldlm_lock, l_res_link);
1013
1014                 if (lock == old_lock)
1015                         break;
1016
1017                 /* llite sometimes wants to match locks that will be
1018                  * canceled when their users drop, but we allow it to match
1019                  * if it passes in CBPENDING and the lock still has users.
1020                  * this is generally only going to be used by children
1021                  * whose parents already hold a lock so forward progress
1022                  * can still happen. */
1023                 if (lock->l_flags & LDLM_FL_CBPENDING &&
1024                     !(flags & LDLM_FL_CBPENDING))
1025                         continue;
1026                 if (!unref && lock->l_flags & LDLM_FL_CBPENDING &&
1027                     lock->l_readers == 0 && lock->l_writers == 0)
1028                         continue;
1029
1030                 if (!(lock->l_req_mode & *mode))
1031                         continue;
1032                 match = lock->l_req_mode;
1033
1034                 if (lock->l_resource->lr_type == LDLM_EXTENT &&
1035                     (lock->l_policy_data.l_extent.start >
1036                      policy->l_extent.start ||
1037                      lock->l_policy_data.l_extent.end < policy->l_extent.end))
1038                         continue;
1039
1040                 if (unlikely(match == LCK_GROUP) &&
1041                     lock->l_resource->lr_type == LDLM_EXTENT &&
1042                     lock->l_policy_data.l_extent.gid != policy->l_extent.gid)
1043                         continue;
1044
1045                 /* We match if we have existing lock with same or wider set
1046                    of bits. */
1047                 if (lock->l_resource->lr_type == LDLM_IBITS &&
1048                      ((lock->l_policy_data.l_inodebits.bits &
1049                       policy->l_inodebits.bits) !=
1050                       policy->l_inodebits.bits))
1051                         continue;
1052
1053                 if (!unref &&
1054                     (lock->l_destroyed || (lock->l_flags & LDLM_FL_FAILED)))
1055                         continue;
1056
1057                 if ((flags & LDLM_FL_LOCAL_ONLY) &&
1058                     !(lock->l_flags & LDLM_FL_LOCAL))
1059                         continue;
1060
1061                 if (flags & LDLM_FL_TEST_LOCK) {
1062                         LDLM_LOCK_GET(lock);
1063                         ldlm_lock_touch_in_lru(lock);
1064                 } else {
1065                         ldlm_lock_addref_internal_nolock(lock, match);
1066                 }
1067                 *mode = match;
1068                 return lock;
1069         }
1070
1071         return NULL;
1072 }
1073
1074 void ldlm_lock_allow_match_locked(struct ldlm_lock *lock)
1075 {
1076         lock->l_flags |= LDLM_FL_LVB_READY;
1077         cfs_waitq_signal(&lock->l_waitq);
1078 }
1079
1080 void ldlm_lock_allow_match(struct ldlm_lock *lock)
1081 {
1082         lock_res_and_lock(lock);
1083         ldlm_lock_allow_match_locked(lock);
1084         unlock_res_and_lock(lock);
1085 }
1086
1087 /* Can be called in two ways:
1088  *
1089  * If 'ns' is NULL, then lockh describes an existing lock that we want to look
1090  * for a duplicate of.
1091  *
1092  * Otherwise, all of the fields must be filled in, to match against.
1093  *
1094  * If 'flags' contains LDLM_FL_LOCAL_ONLY, then only match local locks on the
1095  *     server (ie, connh is NULL)
1096  * If 'flags' contains LDLM_FL_BLOCK_GRANTED, then only locks on the granted
1097  *     list will be considered
1098  * If 'flags' contains LDLM_FL_CBPENDING, then locks that have been marked
1099  *     to be canceled can still be matched as long as they still have reader
1100  *     or writer refernces
1101  * If 'flags' contains LDLM_FL_TEST_LOCK, then don't actually reference a lock,
1102  *     just tell us if we would have matched.
1103  *
1104  * Returns 1 if it finds an already-existing lock that is compatible; in this
1105  * case, lockh is filled in with a addref()ed lock
1106  *
1107  * we also check security context, if that failed we simply return 0 (to keep
1108  * caller code unchanged), the context failure will be discovered by caller
1109  * sometime later.
1110  */
1111 ldlm_mode_t ldlm_lock_match(struct ldlm_namespace *ns, int flags,
1112                             const struct ldlm_res_id *res_id, ldlm_type_t type,
1113                             ldlm_policy_data_t *policy, ldlm_mode_t mode,
1114                             struct lustre_handle *lockh, int unref)
1115 {
1116         struct ldlm_resource *res;
1117         struct ldlm_lock *lock, *old_lock = NULL;
1118         int rc = 0;
1119         ENTRY;
1120
1121         if (ns == NULL) {
1122                 old_lock = ldlm_handle2lock(lockh);
1123                 LASSERT(old_lock);
1124
1125                 ns = ldlm_lock_to_ns(old_lock);
1126                 res_id = &old_lock->l_resource->lr_name;
1127                 type = old_lock->l_resource->lr_type;
1128                 mode = old_lock->l_req_mode;
1129         }
1130
1131         res = ldlm_resource_get(ns, NULL, res_id, type, 0);
1132         if (res == NULL) {
1133                 LASSERT(old_lock == NULL);
1134                 RETURN(0);
1135         }
1136
1137         LDLM_RESOURCE_ADDREF(res);
1138         lock_res(res);
1139
1140         lock = search_queue(&res->lr_granted, &mode, policy, old_lock,
1141                             flags, unref);
1142         if (lock != NULL)
1143                 GOTO(out, rc = 1);
1144         if (flags & LDLM_FL_BLOCK_GRANTED)
1145                 GOTO(out, rc = 0);
1146         lock = search_queue(&res->lr_converting, &mode, policy, old_lock,
1147                             flags, unref);
1148         if (lock != NULL)
1149                 GOTO(out, rc = 1);
1150         lock = search_queue(&res->lr_waiting, &mode, policy, old_lock,
1151                             flags, unref);
1152         if (lock != NULL)
1153                 GOTO(out, rc = 1);
1154
1155         EXIT;
1156  out:
1157         unlock_res(res);
1158         LDLM_RESOURCE_DELREF(res);
1159         ldlm_resource_putref(res);
1160
1161         if (lock) {
1162                 ldlm_lock2handle(lock, lockh);
1163                 if ((flags & LDLM_FL_LVB_READY) &&
1164                     (!(lock->l_flags & LDLM_FL_LVB_READY))) {
1165                         struct l_wait_info lwi;
1166                         if (lock->l_completion_ast) {
1167                                 int err = lock->l_completion_ast(lock,
1168                                                           LDLM_FL_WAIT_NOREPROC,
1169                                                                  NULL);
1170                                 if (err) {
1171                                         if (flags & LDLM_FL_TEST_LOCK)
1172                                                 LDLM_LOCK_RELEASE(lock);
1173                                         else
1174                                                 ldlm_lock_decref_internal(lock,
1175                                                                           mode);
1176                                         rc = 0;
1177                                         goto out2;
1178                                 }
1179                         }
1180
1181                         lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(obd_timeout),
1182                                                NULL, LWI_ON_SIGNAL_NOOP, NULL);
1183
1184                         /* XXX FIXME see comment on CAN_MATCH in lustre_dlm.h */
1185                         l_wait_event(lock->l_waitq,
1186                                      (lock->l_flags & LDLM_FL_LVB_READY), &lwi);
1187                 }
1188         }
1189  out2:
1190         if (rc) {
1191                 LDLM_DEBUG(lock, "matched ("LPU64" "LPU64")",
1192                            (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1193                                 res_id->name[2] : policy->l_extent.start,
1194                            (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1195                                 res_id->name[3] : policy->l_extent.end);
1196
1197                 /* check user's security context */
1198                 if (lock->l_conn_export &&
1199                     sptlrpc_import_check_ctx(
1200                                 class_exp2cliimp(lock->l_conn_export))) {
1201                         if (!(flags & LDLM_FL_TEST_LOCK))
1202                                 ldlm_lock_decref_internal(lock, mode);
1203                         rc = 0;
1204                 }
1205
1206                 if (flags & LDLM_FL_TEST_LOCK)
1207                         LDLM_LOCK_RELEASE(lock);
1208
1209         } else if (!(flags & LDLM_FL_TEST_LOCK)) {/*less verbose for test-only*/
1210                 LDLM_DEBUG_NOLOCK("not matched ns %p type %u mode %u res "
1211                                   LPU64"/"LPU64" ("LPU64" "LPU64")", ns,
1212                                   type, mode, res_id->name[0], res_id->name[1],
1213                                   (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1214                                         res_id->name[2] :policy->l_extent.start,
1215                                   (type == LDLM_PLAIN || type == LDLM_IBITS) ?
1216                                         res_id->name[3] : policy->l_extent.end);
1217         }
1218         if (old_lock)
1219                 LDLM_LOCK_PUT(old_lock);
1220
1221         return rc ? mode : 0;
1222 }
1223
1224 ldlm_mode_t ldlm_revalidate_lock_handle(struct lustre_handle *lockh,
1225                                         __u64 *bits)
1226 {
1227         struct ldlm_lock *lock;
1228         ldlm_mode_t mode = 0;
1229         ENTRY;
1230
1231         lock = ldlm_handle2lock(lockh);
1232         if (lock != NULL) {
1233                 lock_res_and_lock(lock);
1234                 if (lock->l_destroyed || lock->l_flags & LDLM_FL_FAILED)
1235                         GOTO(out, mode);
1236
1237                 if (lock->l_flags & LDLM_FL_CBPENDING &&
1238                     lock->l_readers == 0 && lock->l_writers == 0)
1239                         GOTO(out, mode);
1240
1241                 if (bits)
1242                         *bits = lock->l_policy_data.l_inodebits.bits;
1243                 mode = lock->l_granted_mode;
1244                 ldlm_lock_addref_internal_nolock(lock, mode);
1245         }
1246
1247         EXIT;
1248
1249 out:
1250         if (lock != NULL) {
1251                 unlock_res_and_lock(lock);
1252                 LDLM_LOCK_PUT(lock);
1253         }
1254         return mode;
1255 }
1256 EXPORT_SYMBOL(ldlm_revalidate_lock_handle);
1257
1258 /* Returns a referenced lock */
1259 struct ldlm_lock *ldlm_lock_create(struct ldlm_namespace *ns,
1260                                    const struct ldlm_res_id *res_id,
1261                                    ldlm_type_t type,
1262                                    ldlm_mode_t mode,
1263                                    const struct ldlm_callback_suite *cbs,
1264                                    void *data, __u32 lvb_len)
1265 {
1266         struct ldlm_lock *lock;
1267         struct ldlm_resource *res;
1268         ENTRY;
1269
1270         res = ldlm_resource_get(ns, NULL, res_id, type, 1);
1271         if (res == NULL)
1272                 RETURN(NULL);
1273
1274         lock = ldlm_lock_new(res);
1275
1276         if (lock == NULL)
1277                 RETURN(NULL);
1278
1279         lock->l_req_mode = mode;
1280         lock->l_ast_data = data;
1281         lock->l_pid = cfs_curproc_pid();
1282         lock->l_ns_srv = ns_is_server(ns);
1283         if (cbs) {
1284                 lock->l_blocking_ast = cbs->lcs_blocking;
1285                 lock->l_completion_ast = cbs->lcs_completion;
1286                 lock->l_glimpse_ast = cbs->lcs_glimpse;
1287                 lock->l_weigh_ast = cbs->lcs_weigh;
1288         }
1289
1290         lock->l_tree_node = NULL;
1291         /* if this is the extent lock, allocate the interval tree node */
1292         if (type == LDLM_EXTENT) {
1293                 if (ldlm_interval_alloc(lock) == NULL)
1294                         GOTO(out, 0);
1295         }
1296
1297         if (lvb_len) {
1298                 lock->l_lvb_len = lvb_len;
1299                 OBD_ALLOC(lock->l_lvb_data, lvb_len);
1300                 if (lock->l_lvb_data == NULL)
1301                         GOTO(out, 0);
1302         }
1303
1304         if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_NEW_LOCK))
1305                 GOTO(out, 0);
1306
1307         RETURN(lock);
1308
1309 out:
1310         ldlm_lock_destroy(lock);
1311         LDLM_LOCK_RELEASE(lock);
1312         return NULL;
1313 }
1314
1315 ldlm_error_t ldlm_lock_enqueue(struct ldlm_namespace *ns,
1316                                struct ldlm_lock **lockp,
1317                                void *cookie, int *flags)
1318 {
1319         struct ldlm_lock *lock = *lockp;
1320         struct ldlm_resource *res = lock->l_resource;
1321         int local = ns_is_client(ldlm_res_to_ns(res));
1322         ldlm_processing_policy policy;
1323         ldlm_error_t rc = ELDLM_OK;
1324         struct ldlm_interval *node = NULL;
1325         ENTRY;
1326
1327         lock->l_last_activity = cfs_time_current_sec();
1328         /* policies are not executed on the client or during replay */
1329         if ((*flags & (LDLM_FL_HAS_INTENT|LDLM_FL_REPLAY)) == LDLM_FL_HAS_INTENT
1330             && !local && ns->ns_policy) {
1331                 rc = ns->ns_policy(ns, lockp, cookie, lock->l_req_mode, *flags,
1332                                    NULL);
1333                 if (rc == ELDLM_LOCK_REPLACED) {
1334                         /* The lock that was returned has already been granted,
1335                          * and placed into lockp.  If it's not the same as the
1336                          * one we passed in, then destroy the old one and our
1337                          * work here is done. */
1338                         if (lock != *lockp) {
1339                                 ldlm_lock_destroy(lock);
1340                                 LDLM_LOCK_RELEASE(lock);
1341                         }
1342                         *flags |= LDLM_FL_LOCK_CHANGED;
1343                         RETURN(0);
1344                 } else if (rc != ELDLM_OK ||
1345                            (rc == ELDLM_OK && (*flags & LDLM_FL_INTENT_ONLY))) {
1346                         ldlm_lock_destroy(lock);
1347                         RETURN(rc);
1348                 }
1349         }
1350
1351         /* For a replaying lock, it might be already in granted list. So
1352          * unlinking the lock will cause the interval node to be freed, we
1353          * have to allocate the interval node early otherwise we can't regrant
1354          * this lock in the future. - jay */
1355         if (!local && (*flags & LDLM_FL_REPLAY) && res->lr_type == LDLM_EXTENT)
1356                 OBD_SLAB_ALLOC_PTR_GFP(node, ldlm_interval_slab, CFS_ALLOC_IO);
1357
1358         lock_res_and_lock(lock);
1359         if (local && lock->l_req_mode == lock->l_granted_mode) {
1360                 /* The server returned a blocked lock, but it was granted
1361                  * before we got a chance to actually enqueue it.  We don't
1362                  * need to do anything else. */
1363                 *flags &= ~(LDLM_FL_BLOCK_GRANTED |
1364                             LDLM_FL_BLOCK_CONV | LDLM_FL_BLOCK_WAIT);
1365                 GOTO(out, ELDLM_OK);
1366         }
1367
1368         ldlm_resource_unlink_lock(lock);
1369         if (res->lr_type == LDLM_EXTENT && lock->l_tree_node == NULL) {
1370                 if (node == NULL) {
1371                         ldlm_lock_destroy_nolock(lock);
1372                         GOTO(out, rc = -ENOMEM);
1373                 }
1374
1375                 CFS_INIT_LIST_HEAD(&node->li_group);
1376                 ldlm_interval_attach(node, lock);
1377                 node = NULL;
1378         }
1379
1380         /* Some flags from the enqueue want to make it into the AST, via the
1381          * lock's l_flags. */
1382         lock->l_flags |= *flags & LDLM_AST_DISCARD_DATA;
1383
1384         /* This distinction between local lock trees is very important; a client
1385          * namespace only has information about locks taken by that client, and
1386          * thus doesn't have enough information to decide for itself if it can
1387          * be granted (below).  In this case, we do exactly what the server
1388          * tells us to do, as dictated by the 'flags'.
1389          *
1390          * We do exactly the same thing during recovery, when the server is
1391          * more or less trusting the clients not to lie.
1392          *
1393          * FIXME (bug 268): Detect obvious lies by checking compatibility in
1394          * granted/converting queues. */
1395         if (local) {
1396                 if (*flags & LDLM_FL_BLOCK_CONV)
1397                         ldlm_resource_add_lock(res, &res->lr_converting, lock);
1398                 else if (*flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED))
1399                         ldlm_resource_add_lock(res, &res->lr_waiting, lock);
1400                 else
1401                         ldlm_grant_lock(lock, NULL);
1402                 GOTO(out, ELDLM_OK);
1403         } else if (*flags & LDLM_FL_REPLAY) {
1404                 if (*flags & LDLM_FL_BLOCK_CONV) {
1405                         ldlm_resource_add_lock(res, &res->lr_converting, lock);
1406                         GOTO(out, ELDLM_OK);
1407                 } else if (*flags & LDLM_FL_BLOCK_WAIT) {
1408                         ldlm_resource_add_lock(res, &res->lr_waiting, lock);
1409                         GOTO(out, ELDLM_OK);
1410                 } else if (*flags & LDLM_FL_BLOCK_GRANTED) {
1411                         ldlm_grant_lock(lock, NULL);
1412                         GOTO(out, ELDLM_OK);
1413                 }
1414                 /* If no flags, fall through to normal enqueue path. */
1415         }
1416
1417         policy = ldlm_processing_policy_table[res->lr_type];
1418         policy(lock, flags, 1, &rc, NULL);
1419         GOTO(out, rc);
1420 out:
1421         unlock_res_and_lock(lock);
1422         if (node)
1423                 OBD_SLAB_FREE(node, ldlm_interval_slab, sizeof(*node));
1424         return rc;
1425 }
1426
1427 /* Must be called with namespace taken: queue is waiting or converting. */
1428 int ldlm_reprocess_queue(struct ldlm_resource *res, cfs_list_t *queue,
1429                          cfs_list_t *work_list)
1430 {
1431         cfs_list_t *tmp, *pos;
1432         ldlm_processing_policy policy;
1433         int flags;
1434         int rc = LDLM_ITER_CONTINUE;
1435         ldlm_error_t err;
1436         ENTRY;
1437
1438         check_res_locked(res);
1439
1440         policy = ldlm_processing_policy_table[res->lr_type];
1441         LASSERT(policy);
1442
1443         cfs_list_for_each_safe(tmp, pos, queue) {
1444                 struct ldlm_lock *pending;
1445                 pending = cfs_list_entry(tmp, struct ldlm_lock, l_res_link);
1446
1447                 CDEBUG(D_INFO, "Reprocessing lock %p\n", pending);
1448
1449                 flags = 0;
1450                 rc = policy(pending, &flags, 0, &err, work_list);
1451                 if (rc != LDLM_ITER_CONTINUE)
1452                         break;
1453         }
1454
1455         RETURN(rc);
1456 }
1457
1458 /* Helper function for ldlm_run_ast_work().
1459  *
1460  * Send an existing rpc set specified by @arg->set and then
1461  * destroy it. Create new one if @do_create flag is set. */
1462 static int ldlm_deliver_cb_set(struct ldlm_cb_set_arg *arg, int do_create)
1463 {
1464         int rc = 0;
1465         ENTRY;
1466
1467         if (arg->set) {
1468                 ptlrpc_set_wait(arg->set);
1469                 if (arg->type == LDLM_BL_CALLBACK)
1470                         OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_GLIMPSE, 2);
1471                 ptlrpc_set_destroy(arg->set);
1472                 arg->set = NULL;
1473                 arg->rpcs = 0;
1474         }
1475
1476         if (do_create) {
1477                 arg->set = ptlrpc_prep_set();
1478                 if (arg->set == NULL)
1479                         rc = -ENOMEM;
1480         }
1481
1482         RETURN(rc);
1483 }
1484
1485 static int
1486 ldlm_work_bl_ast_lock(cfs_list_t *tmp, struct ldlm_cb_set_arg *arg)
1487 {
1488         struct ldlm_lock_desc d;
1489         struct ldlm_lock *lock = cfs_list_entry(tmp, struct ldlm_lock,
1490                                                 l_bl_ast);
1491         int rc;
1492         ENTRY;
1493
1494         /* nobody should touch l_bl_ast */
1495         lock_res_and_lock(lock);
1496         cfs_list_del_init(&lock->l_bl_ast);
1497
1498         LASSERT(lock->l_flags & LDLM_FL_AST_SENT);
1499         LASSERT(lock->l_bl_ast_run == 0);
1500         LASSERT(lock->l_blocking_lock);
1501         lock->l_bl_ast_run++;
1502         unlock_res_and_lock(lock);
1503
1504         ldlm_lock2desc(lock->l_blocking_lock, &d);
1505
1506         rc = lock->l_blocking_ast(lock, &d, (void *)arg,
1507                                   LDLM_CB_BLOCKING);
1508         LDLM_LOCK_RELEASE(lock->l_blocking_lock);
1509         lock->l_blocking_lock = NULL;
1510         LDLM_LOCK_RELEASE(lock);
1511
1512         RETURN(rc);
1513 }
1514
1515 static int
1516 ldlm_work_cp_ast_lock(cfs_list_t *tmp, struct ldlm_cb_set_arg *arg)
1517 {
1518         struct ldlm_lock *lock = cfs_list_entry(tmp, struct ldlm_lock, l_cp_ast);
1519         ldlm_completion_callback completion_callback;
1520         int rc = 0;
1521         ENTRY;
1522
1523         /* It's possible to receive a completion AST before we've set
1524          * the l_completion_ast pointer: either because the AST arrived
1525          * before the reply, or simply because there's a small race
1526          * window between receiving the reply and finishing the local
1527          * enqueue. (bug 842)
1528          *
1529          * This can't happen with the blocking_ast, however, because we
1530          * will never call the local blocking_ast until we drop our
1531          * reader/writer reference, which we won't do until we get the
1532          * reply and finish enqueueing. */
1533
1534         /* nobody should touch l_cp_ast */
1535         lock_res_and_lock(lock);
1536         cfs_list_del_init(&lock->l_cp_ast);
1537         LASSERT(lock->l_flags & LDLM_FL_CP_REQD);
1538         /* save l_completion_ast since it can be changed by
1539          * mds_intent_policy(), see bug 14225 */
1540         completion_callback = lock->l_completion_ast;
1541         lock->l_flags &= ~LDLM_FL_CP_REQD;
1542         unlock_res_and_lock(lock);
1543
1544         if (completion_callback != NULL)
1545                 rc = completion_callback(lock, 0, (void *)arg);
1546         LDLM_LOCK_RELEASE(lock);
1547
1548         RETURN(rc);
1549 }
1550
1551 static int
1552 ldlm_work_revoke_ast_lock(cfs_list_t *tmp, struct ldlm_cb_set_arg *arg)
1553 {
1554         struct ldlm_lock_desc desc;
1555         struct ldlm_lock *lock = cfs_list_entry(tmp, struct ldlm_lock,
1556                                                 l_rk_ast);
1557         int rc;
1558         ENTRY;
1559
1560         cfs_list_del_init(&lock->l_rk_ast);
1561
1562         /* the desc just pretend to exclusive */
1563         ldlm_lock2desc(lock, &desc);
1564         desc.l_req_mode = LCK_EX;
1565         desc.l_granted_mode = 0;
1566
1567         rc = lock->l_blocking_ast(lock, &desc, (void*)arg, LDLM_CB_BLOCKING);
1568         LDLM_LOCK_RELEASE(lock);
1569
1570         RETURN(rc);
1571 }
1572
1573 int ldlm_run_ast_work(struct ldlm_namespace *ns, cfs_list_t *rpc_list,
1574                       ldlm_desc_ast_t ast_type)
1575 {
1576         struct ldlm_cb_set_arg arg = { 0 };
1577         cfs_list_t *tmp, *pos;
1578         int (*work_ast_lock)(cfs_list_t *tmp, struct ldlm_cb_set_arg *arg);
1579         unsigned int max_ast_count;
1580         int rc;
1581         ENTRY;
1582
1583         if (cfs_list_empty(rpc_list))
1584                 RETURN(0);
1585
1586         rc = ldlm_deliver_cb_set(&arg, 1);
1587         if (rc != 0)
1588                 RETURN(rc);
1589
1590         switch (ast_type) {
1591         case LDLM_WORK_BL_AST:
1592                 arg.type = LDLM_BL_CALLBACK;
1593                 work_ast_lock = ldlm_work_bl_ast_lock;
1594                 break;
1595         case LDLM_WORK_CP_AST:
1596                 arg.type = LDLM_CP_CALLBACK;
1597                 work_ast_lock = ldlm_work_cp_ast_lock;
1598                 break;
1599         case LDLM_WORK_REVOKE_AST:
1600                 arg.type = LDLM_BL_CALLBACK;
1601                 work_ast_lock = ldlm_work_revoke_ast_lock;
1602                 break;
1603         default:
1604                 LBUG();
1605         }
1606
1607         max_ast_count = ns->ns_max_parallel_ast ? : UINT_MAX;
1608
1609         cfs_list_for_each_safe(tmp, pos, rpc_list) {
1610                 (void)work_ast_lock(tmp, &arg);
1611                 if (arg.rpcs > max_ast_count) {
1612                         rc = ldlm_deliver_cb_set(&arg, 1);
1613                         if (rc != 0)
1614                                 break;
1615                 }
1616         }
1617
1618         (void)ldlm_deliver_cb_set(&arg, 0);
1619
1620         if (rc == 0 && cfs_atomic_read(&arg.restart))
1621                 rc = -ERESTART;
1622
1623         RETURN(rc);
1624 }
1625
1626 static int reprocess_one_queue(struct ldlm_resource *res, void *closure)
1627 {
1628         ldlm_reprocess_all(res);
1629         return LDLM_ITER_CONTINUE;
1630 }
1631
1632 static int ldlm_reprocess_res(cfs_hash_t *hs, cfs_hash_bd_t *bd,
1633                               cfs_hlist_node_t *hnode, void *arg)
1634 {
1635         struct ldlm_resource *res = cfs_hash_object(hs, hnode);
1636         int    rc;
1637
1638         rc = reprocess_one_queue(res, arg);
1639
1640         return rc == LDLM_ITER_STOP;
1641 }
1642
1643 void ldlm_reprocess_all_ns(struct ldlm_namespace *ns)
1644 {
1645         ENTRY;
1646
1647         if (ns != NULL) {
1648                 cfs_hash_for_each_nolock(ns->ns_rs_hash,
1649                                          ldlm_reprocess_res, NULL);
1650         }
1651         EXIT;
1652 }
1653
1654 void ldlm_reprocess_all(struct ldlm_resource *res)
1655 {
1656         CFS_LIST_HEAD(rpc_list);
1657         int rc;
1658         ENTRY;
1659
1660         /* Local lock trees don't get reprocessed. */
1661         if (ns_is_client(ldlm_res_to_ns(res))) {
1662                 EXIT;
1663                 return;
1664         }
1665
1666  restart:
1667         lock_res(res);
1668         rc = ldlm_reprocess_queue(res, &res->lr_converting, &rpc_list);
1669         if (rc == LDLM_ITER_CONTINUE)
1670                 ldlm_reprocess_queue(res, &res->lr_waiting, &rpc_list);
1671         unlock_res(res);
1672
1673         rc = ldlm_run_ast_work(ldlm_res_to_ns(res), &rpc_list,
1674                                LDLM_WORK_CP_AST);
1675         if (rc == -ERESTART) {
1676                 LASSERT(cfs_list_empty(&rpc_list));
1677                 goto restart;
1678         }
1679         EXIT;
1680 }
1681
1682 void ldlm_cancel_callback(struct ldlm_lock *lock)
1683 {
1684         check_res_locked(lock->l_resource);
1685         if (!(lock->l_flags & LDLM_FL_CANCEL)) {
1686                 lock->l_flags |= LDLM_FL_CANCEL;
1687                 if (lock->l_blocking_ast) {
1688                         // l_check_no_ns_lock(ns);
1689                         unlock_res_and_lock(lock);
1690                         lock->l_blocking_ast(lock, NULL, lock->l_ast_data,
1691                                              LDLM_CB_CANCELING);
1692                         lock_res_and_lock(lock);
1693                 } else {
1694                         LDLM_DEBUG(lock, "no blocking ast");
1695                 }
1696         }
1697         lock->l_flags |= LDLM_FL_BL_DONE;
1698 }
1699
1700 void ldlm_unlink_lock_skiplist(struct ldlm_lock *req)
1701 {
1702         if (req->l_resource->lr_type != LDLM_PLAIN &&
1703             req->l_resource->lr_type != LDLM_IBITS)
1704                 return;
1705
1706         cfs_list_del_init(&req->l_sl_policy);
1707         cfs_list_del_init(&req->l_sl_mode);
1708 }
1709
1710 void ldlm_lock_cancel(struct ldlm_lock *lock)
1711 {
1712         struct ldlm_resource *res;
1713         struct ldlm_namespace *ns;
1714         ENTRY;
1715
1716         lock_res_and_lock(lock);
1717
1718         res = lock->l_resource;
1719         ns  = ldlm_res_to_ns(res);
1720
1721         /* Please do not, no matter how tempting, remove this LBUG without
1722          * talking to me first. -phik */
1723         if (lock->l_readers || lock->l_writers) {
1724                 LDLM_ERROR(lock, "lock still has references");
1725                 LBUG();
1726         }
1727
1728         ldlm_del_waiting_lock(lock);
1729
1730         /* Releases cancel callback. */
1731         ldlm_cancel_callback(lock);
1732
1733         /* Yes, second time, just in case it was added again while we were
1734            running with no res lock in ldlm_cancel_callback */
1735         ldlm_del_waiting_lock(lock);
1736         ldlm_resource_unlink_lock(lock);
1737         ldlm_lock_destroy_nolock(lock);
1738
1739         if (lock->l_granted_mode == lock->l_req_mode)
1740                 ldlm_pool_del(&ns->ns_pool, lock);
1741
1742         /* Make sure we will not be called again for same lock what is possible
1743          * if not to zero out lock->l_granted_mode */
1744         lock->l_granted_mode = LCK_MINMODE;
1745         unlock_res_and_lock(lock);
1746
1747         EXIT;
1748 }
1749
1750 int ldlm_lock_set_data(struct lustre_handle *lockh, void *data)
1751 {
1752         struct ldlm_lock *lock = ldlm_handle2lock(lockh);
1753         int rc = -EINVAL;
1754         ENTRY;
1755
1756         if (lock) {
1757                 if (lock->l_ast_data == NULL)
1758                         lock->l_ast_data = data;
1759                 if (lock->l_ast_data == data)
1760                         rc = 0;
1761                 LDLM_LOCK_PUT(lock);
1762         }
1763         RETURN(rc);
1764 }
1765 EXPORT_SYMBOL(ldlm_lock_set_data);
1766
1767 int ldlm_cancel_locks_for_export_cb(cfs_hash_t *hs, cfs_hash_bd_t *bd,
1768                                     cfs_hlist_node_t *hnode, void *data)
1769
1770 {
1771         struct obd_export    *exp  = data;
1772         struct ldlm_lock     *lock = cfs_hash_object(hs, hnode);
1773         struct ldlm_resource *res;
1774
1775         res = ldlm_resource_getref(lock->l_resource);
1776         LDLM_LOCK_GET(lock);
1777
1778         LDLM_DEBUG(lock, "export %p", exp);
1779         ldlm_res_lvbo_update(res, NULL, 1);
1780         ldlm_lock_cancel(lock);
1781         ldlm_reprocess_all(res);
1782         ldlm_resource_putref(res);
1783         LDLM_LOCK_RELEASE(lock);
1784         return 0;
1785 }
1786
1787 void ldlm_cancel_locks_for_export(struct obd_export *exp)
1788 {
1789         cfs_hash_for_each_empty(exp->exp_lock_hash,
1790                                 ldlm_cancel_locks_for_export_cb, exp);
1791 }
1792
1793 /**
1794  * Downgrade an exclusive lock.
1795  *
1796  * A fast variant of ldlm_lock_convert for convertion of exclusive
1797  * locks. The convertion is always successful.
1798  *
1799  * \param lock A lock to convert
1800  * \param new_mode new lock mode
1801  */
1802 void ldlm_lock_downgrade(struct ldlm_lock *lock, int new_mode)
1803 {
1804         ENTRY;
1805
1806         LASSERT(lock->l_granted_mode & (LCK_PW | LCK_EX));
1807         LASSERT(new_mode == LCK_COS);
1808
1809         lock_res_and_lock(lock);
1810         ldlm_resource_unlink_lock(lock);
1811         /*
1812          * Remove the lock from pool as it will be added again in
1813          * ldlm_grant_lock() called below.
1814          */
1815         ldlm_pool_del(&ldlm_lock_to_ns(lock)->ns_pool, lock);
1816
1817         lock->l_req_mode = new_mode;
1818         ldlm_grant_lock(lock, NULL);
1819         unlock_res_and_lock(lock);
1820         ldlm_reprocess_all(lock->l_resource);
1821
1822         EXIT;
1823 }
1824
1825 struct ldlm_resource *ldlm_lock_convert(struct ldlm_lock *lock, int new_mode,
1826                                         __u32 *flags)
1827 {
1828         CFS_LIST_HEAD(rpc_list);
1829         struct ldlm_resource *res;
1830         struct ldlm_namespace *ns;
1831         int granted = 0;
1832         int old_mode, rc;
1833         struct sl_insert_point prev;
1834         ldlm_error_t err;
1835         struct ldlm_interval *node;
1836         ENTRY;
1837
1838         if (new_mode == lock->l_granted_mode) { // No changes? Just return.
1839                 *flags |= LDLM_FL_BLOCK_GRANTED;
1840                 RETURN(lock->l_resource);
1841         }
1842
1843         /* I can't check the type of lock here because the bitlock of lock
1844          * is not held here, so do the allocation blindly. -jay */
1845         OBD_SLAB_ALLOC_PTR_GFP(node, ldlm_interval_slab, CFS_ALLOC_IO);
1846         if (node == NULL)  /* Actually, this causes EDEADLOCK to be returned */
1847                 RETURN(NULL);
1848
1849         LASSERTF((new_mode == LCK_PW && lock->l_granted_mode == LCK_PR),
1850                  "new_mode %u, granted %u\n", new_mode, lock->l_granted_mode);
1851
1852         lock_res_and_lock(lock);
1853
1854         res = lock->l_resource;
1855         ns  = ldlm_res_to_ns(res);
1856
1857         old_mode = lock->l_req_mode;
1858         lock->l_req_mode = new_mode;
1859         if (res->lr_type == LDLM_PLAIN || res->lr_type == LDLM_IBITS) {
1860                 /* remember the lock position where the lock might be
1861                  * added back to the granted list later and also
1862                  * remember the join mode for skiplist fixing. */
1863                 prev.res_link = lock->l_res_link.prev;
1864                 prev.mode_link = lock->l_sl_mode.prev;
1865                 prev.policy_link = lock->l_sl_policy.prev;
1866                 ldlm_resource_unlink_lock(lock);
1867         } else {
1868                 ldlm_resource_unlink_lock(lock);
1869                 if (res->lr_type == LDLM_EXTENT) {
1870                         /* FIXME: ugly code, I have to attach the lock to a
1871                          * interval node again since perhaps it will be granted
1872                          * soon */
1873                         CFS_INIT_LIST_HEAD(&node->li_group);
1874                         ldlm_interval_attach(node, lock);
1875                         node = NULL;
1876                 }
1877         }
1878
1879         /*
1880          * Remove old lock from the pool before adding the lock with new
1881          * mode below in ->policy()
1882          */
1883         ldlm_pool_del(&ns->ns_pool, lock);
1884
1885         /* If this is a local resource, put it on the appropriate list. */
1886         if (ns_is_client(ldlm_res_to_ns(res))) {
1887                 if (*flags & (LDLM_FL_BLOCK_CONV | LDLM_FL_BLOCK_GRANTED)) {
1888                         ldlm_resource_add_lock(res, &res->lr_converting, lock);
1889                 } else {
1890                         /* This should never happen, because of the way the
1891                          * server handles conversions. */
1892                         LDLM_ERROR(lock, "Erroneous flags %d on local lock\n",
1893                                    *flags);
1894                         LBUG();
1895
1896                         ldlm_grant_lock(lock, &rpc_list);
1897                         granted = 1;
1898                         /* FIXME: completion handling not with lr_lock held ! */
1899                         if (lock->l_completion_ast)
1900                                 lock->l_completion_ast(lock, 0, NULL);
1901                 }
1902         } else {
1903                 int pflags = 0;
1904                 ldlm_processing_policy policy;
1905                 policy = ldlm_processing_policy_table[res->lr_type];
1906                 rc = policy(lock, &pflags, 0, &err, &rpc_list);
1907                 if (rc == LDLM_ITER_STOP) {
1908                         lock->l_req_mode = old_mode;
1909                         if (res->lr_type == LDLM_EXTENT)
1910                                 ldlm_extent_add_lock(res, lock);
1911                         else
1912                                 ldlm_granted_list_add_lock(lock, &prev);
1913
1914                         res = NULL;
1915                 } else {
1916                         *flags |= LDLM_FL_BLOCK_GRANTED;
1917                         granted = 1;
1918                 }
1919         }
1920         unlock_res_and_lock(lock);
1921
1922         if (granted)
1923                 ldlm_run_ast_work(ns, &rpc_list, LDLM_WORK_CP_AST);
1924         if (node)
1925                 OBD_SLAB_FREE(node, ldlm_interval_slab, sizeof(*node));
1926         RETURN(res);
1927 }
1928
1929 void ldlm_lock_dump(int level, struct ldlm_lock *lock, int pos)
1930 {
1931         struct obd_device *obd = NULL;
1932
1933         if (!((libcfs_debug | D_ERROR) & level))
1934                 return;
1935
1936         if (!lock) {
1937                 CDEBUG(level, "  NULL LDLM lock\n");
1938                 return;
1939         }
1940
1941         CDEBUG(level," -- Lock dump: %p/"LPX64" (rc: %d) (pos: %d) (pid: %d)\n",
1942                lock, lock->l_handle.h_cookie, cfs_atomic_read(&lock->l_refc),
1943                pos, lock->l_pid);
1944         if (lock->l_conn_export != NULL)
1945                 obd = lock->l_conn_export->exp_obd;
1946         if (lock->l_export && lock->l_export->exp_connection) {
1947                 CDEBUG(level, "  Node: NID %s (rhandle: "LPX64")\n",
1948                      libcfs_nid2str(lock->l_export->exp_connection->c_peer.nid),
1949                      lock->l_remote_handle.cookie);
1950         } else if (obd == NULL) {
1951                 CDEBUG(level, "  Node: local\n");
1952         } else {
1953                 struct obd_import *imp = obd->u.cli.cl_import;
1954                 CDEBUG(level, "  Node: NID %s (rhandle: "LPX64")\n",
1955                        libcfs_nid2str(imp->imp_connection->c_peer.nid),
1956                        lock->l_remote_handle.cookie);
1957         }
1958         CDEBUG(level, "  Resource: %p ("LPU64"/"LPU64"/"LPU64")\n",
1959                   lock->l_resource,
1960                   lock->l_resource->lr_name.name[0],
1961                   lock->l_resource->lr_name.name[1],
1962                   lock->l_resource->lr_name.name[2]);
1963         CDEBUG(level, "  Req mode: %s, grant mode: %s, rc: %u, read: %d, "
1964                "write: %d flags: "LPX64"\n", ldlm_lockname[lock->l_req_mode],
1965                ldlm_lockname[lock->l_granted_mode],
1966                cfs_atomic_read(&lock->l_refc), lock->l_readers, lock->l_writers,
1967                lock->l_flags);
1968         if (lock->l_resource->lr_type == LDLM_EXTENT)
1969                 CDEBUG(level, "  Extent: "LPU64" -> "LPU64
1970                        " (req "LPU64"-"LPU64")\n",
1971                        lock->l_policy_data.l_extent.start,
1972                        lock->l_policy_data.l_extent.end,
1973                        lock->l_req_extent.start, lock->l_req_extent.end);
1974         else if (lock->l_resource->lr_type == LDLM_FLOCK)
1975                 CDEBUG(level, "  Pid: %d Extent: "LPU64" -> "LPU64"\n",
1976                        lock->l_policy_data.l_flock.pid,
1977                        lock->l_policy_data.l_flock.start,
1978                        lock->l_policy_data.l_flock.end);
1979        else if (lock->l_resource->lr_type == LDLM_IBITS)
1980                 CDEBUG(level, "  Bits: "LPX64"\n",
1981                        lock->l_policy_data.l_inodebits.bits);
1982 }
1983
1984 void ldlm_lock_dump_handle(int level, struct lustre_handle *lockh)
1985 {
1986         struct ldlm_lock *lock;
1987
1988         if (!((libcfs_debug | D_ERROR) & level))
1989                 return;
1990
1991         lock = ldlm_handle2lock(lockh);
1992         if (lock == NULL)
1993                 return;
1994
1995         ldlm_lock_dump(D_OTHER, lock, 0);
1996
1997         LDLM_LOCK_PUT(lock);
1998 }
1999
2000 void _ldlm_lock_debug(struct ldlm_lock *lock, __u32 level,
2001                       struct libcfs_debug_msg_data *data, const char *fmt,
2002                       ...)
2003 {
2004         va_list args;
2005         cfs_debug_limit_state_t *cdls = data->msg_cdls;
2006
2007         va_start(args, fmt);
2008
2009         if (lock->l_resource == NULL) {
2010                 libcfs_debug_vmsg2(cdls, data->msg_subsys, level,data->msg_file,
2011                                    data->msg_fn, data->msg_line, fmt, args,
2012                        " ns: \?\? lock: %p/"LPX64" lrc: %d/%d,%d mode: %s/%s "
2013                        "res: \?\? rrc=\?\? type: \?\?\? flags: "LPX64" remote: "
2014                        LPX64" expref: %d pid: %u timeout: %lu\n", lock,
2015                        lock->l_handle.h_cookie, cfs_atomic_read(&lock->l_refc),
2016                        lock->l_readers, lock->l_writers,
2017                        ldlm_lockname[lock->l_granted_mode],
2018                        ldlm_lockname[lock->l_req_mode],
2019                        lock->l_flags, lock->l_remote_handle.cookie,
2020                        lock->l_export ?
2021                        cfs_atomic_read(&lock->l_export->exp_refcount) : -99,
2022                        lock->l_pid, lock->l_callback_timeout);
2023                 va_end(args);
2024                 return;
2025         }
2026
2027         switch (lock->l_resource->lr_type) {
2028         case LDLM_EXTENT:
2029                 libcfs_debug_vmsg2(cdls, data->msg_subsys, level,data->msg_file,
2030                                    data->msg_fn, data->msg_line, fmt, args,
2031                        " ns: %s lock: %p/"LPX64" lrc: %d/%d,%d mode: %s/%s "
2032                        "res: "LPU64"/"LPU64" rrc: %d type: %s ["LPU64"->"LPU64
2033                        "] (req "LPU64"->"LPU64") flags: "LPX64" remote: "LPX64
2034                        " expref: %d pid: %u timeout %lu\n",
2035                        ldlm_lock_to_ns_name(lock), lock,
2036                        lock->l_handle.h_cookie, cfs_atomic_read(&lock->l_refc),
2037                        lock->l_readers, lock->l_writers,
2038                        ldlm_lockname[lock->l_granted_mode],
2039                        ldlm_lockname[lock->l_req_mode],
2040                        lock->l_resource->lr_name.name[0],
2041                        lock->l_resource->lr_name.name[1],
2042                        cfs_atomic_read(&lock->l_resource->lr_refcount),
2043                        ldlm_typename[lock->l_resource->lr_type],
2044                        lock->l_policy_data.l_extent.start,
2045                        lock->l_policy_data.l_extent.end,
2046                        lock->l_req_extent.start, lock->l_req_extent.end,
2047                        lock->l_flags, lock->l_remote_handle.cookie,
2048                        lock->l_export ?
2049                        cfs_atomic_read(&lock->l_export->exp_refcount) : -99,
2050                        lock->l_pid, lock->l_callback_timeout);
2051                 break;
2052
2053         case LDLM_FLOCK:
2054                 libcfs_debug_vmsg2(cdls, data->msg_subsys, level,data->msg_file,
2055                                    data->msg_fn, data->msg_line, fmt, args,
2056                        " ns: %s lock: %p/"LPX64" lrc: %d/%d,%d mode: %s/%s "
2057                        "res: "LPU64"/"LPU64" rrc: %d type: %s pid: %d "
2058                        "["LPU64"->"LPU64"] flags: "LPX64" remote: "LPX64
2059                        " expref: %d pid: %u timeout: %lu\n",
2060                        ldlm_lock_to_ns_name(lock), lock,
2061                        lock->l_handle.h_cookie, cfs_atomic_read(&lock->l_refc),
2062                        lock->l_readers, lock->l_writers,
2063                        ldlm_lockname[lock->l_granted_mode],
2064                        ldlm_lockname[lock->l_req_mode],
2065                        lock->l_resource->lr_name.name[0],
2066                        lock->l_resource->lr_name.name[1],
2067                        cfs_atomic_read(&lock->l_resource->lr_refcount),
2068                        ldlm_typename[lock->l_resource->lr_type],
2069                        lock->l_policy_data.l_flock.pid,
2070                        lock->l_policy_data.l_flock.start,
2071                        lock->l_policy_data.l_flock.end,
2072                        lock->l_flags, lock->l_remote_handle.cookie,
2073                        lock->l_export ?
2074                        cfs_atomic_read(&lock->l_export->exp_refcount) : -99,
2075                        lock->l_pid, lock->l_callback_timeout);
2076                 break;
2077
2078         case LDLM_IBITS:
2079                 libcfs_debug_vmsg2(cdls, data->msg_subsys, level,data->msg_file,
2080                                    data->msg_fn, data->msg_line, fmt, args,
2081                        " ns: %s lock: %p/"LPX64" lrc: %d/%d,%d mode: %s/%s "
2082                        "res: "LPU64"/"LPU64" bits "LPX64" rrc: %d type: %s "
2083                        "flags: "LPX64" remote: "LPX64" expref: %d "
2084                        "pid: %u timeout: %lu\n",
2085                        ldlm_lock_to_ns_name(lock),
2086                        lock, lock->l_handle.h_cookie,
2087                        cfs_atomic_read (&lock->l_refc),
2088                        lock->l_readers, lock->l_writers,
2089                        ldlm_lockname[lock->l_granted_mode],
2090                        ldlm_lockname[lock->l_req_mode],
2091                        lock->l_resource->lr_name.name[0],
2092                        lock->l_resource->lr_name.name[1],
2093                        lock->l_policy_data.l_inodebits.bits,
2094                        cfs_atomic_read(&lock->l_resource->lr_refcount),
2095                        ldlm_typename[lock->l_resource->lr_type],
2096                        lock->l_flags, lock->l_remote_handle.cookie,
2097                        lock->l_export ?
2098                        cfs_atomic_read(&lock->l_export->exp_refcount) : -99,
2099                        lock->l_pid, lock->l_callback_timeout);
2100                 break;
2101
2102         default:
2103                 libcfs_debug_vmsg2(cdls, data->msg_subsys, level,data->msg_file,
2104                                    data->msg_fn, data->msg_line, fmt, args,
2105                        " ns: %s lock: %p/"LPX64" lrc: %d/%d,%d mode: %s/%s "
2106                        "res: "LPU64"/"LPU64" rrc: %d type: %s flags: "LPX64" "
2107                        "remote: "LPX64" expref: %d pid: %u timeout %lu\n",
2108                        ldlm_lock_to_ns_name(lock),
2109                        lock, lock->l_handle.h_cookie,
2110                        cfs_atomic_read (&lock->l_refc),
2111                        lock->l_readers, lock->l_writers,
2112                        ldlm_lockname[lock->l_granted_mode],
2113                        ldlm_lockname[lock->l_req_mode],
2114                        lock->l_resource->lr_name.name[0],
2115                        lock->l_resource->lr_name.name[1],
2116                        cfs_atomic_read(&lock->l_resource->lr_refcount),
2117                        ldlm_typename[lock->l_resource->lr_type],
2118                        lock->l_flags, lock->l_remote_handle.cookie,
2119                        lock->l_export ?
2120                        cfs_atomic_read(&lock->l_export->exp_refcount) : -99,
2121                        lock->l_pid, lock->l_callback_timeout);
2122                 break;
2123         }
2124         va_end(args);
2125 }
2126 EXPORT_SYMBOL(_ldlm_lock_debug);