Whamcloud - gitweb
Land b1_2 onto HEAD (20040304_171022)
[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  *  Copyright (c) 2002, 2003 Cluster File Systems, Inc.
5  *   Author: Peter Braam <braam@clusterfs.com>
6  *   Author: Phil Schwan <phil@clusterfs.com>
7  *
8  *   This file is part of Lustre, http://www.lustre.org.
9  *
10  *   Lustre is free software; you can redistribute it and/or
11  *   modify it under the terms of version 2 of the GNU General Public
12  *   License as published by the Free Software Foundation.
13  *
14  *   Lustre is distributed in the hope that it will be useful,
15  *   but WITHOUT ANY WARRANTY; without even the implied warranty of
16  *   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
17  *   GNU General Public License for more details.
18  *
19  *   You should have received a copy of the GNU General Public License
20  *   along with Lustre; if not, write to the Free Software
21  *   Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
22  */
23
24 #define DEBUG_SUBSYSTEM S_LDLM
25
26 #ifdef __KERNEL__
27 # include <linux/slab.h>
28 # include <linux/module.h>
29 # include <linux/lustre_dlm.h>
30 #else
31 # include <liblustre.h>
32 # include <linux/kp30.h>
33 #endif
34
35 #include <linux/obd_class.h>
36 #include "ldlm_internal.h"
37
38 //struct lustre_lock ldlm_everything_lock;
39
40 /* lock types */
41 char *ldlm_lockname[] = {
42         [0] "--",
43         [LCK_EX] "EX",
44         [LCK_PW] "PW",
45         [LCK_PR] "PR",
46         [LCK_CW] "CW",
47         [LCK_CR] "CR",
48         [LCK_NL] "NL"
49 };
50 char *ldlm_typename[] = {
51         [LDLM_PLAIN] "PLN",
52         [LDLM_EXTENT] "EXT",
53         [LDLM_FLOCK] "FLK",
54 };
55
56 char *ldlm_it2str(int it)
57 {
58         switch (it) {
59         case IT_OPEN:
60                 return "open";
61         case IT_CREAT:
62                 return "creat";
63         case (IT_OPEN | IT_CREAT):
64                 return "open|creat";
65         case IT_READDIR:
66                 return "readdir";
67         case IT_GETATTR:
68                 return "getattr";
69         case IT_LOOKUP:
70                 return "lookup";
71         case IT_UNLINK:
72                 return "unlink";
73         case IT_GETXATTR:
74                 return "getxattr";
75         default:
76                 CERROR("Unknown intent %d\n", it);
77                 return "UNKNOWN";
78         }
79 }
80
81 extern kmem_cache_t *ldlm_lock_slab;
82 struct lustre_lock ldlm_handle_lock;
83
84 static ldlm_processing_policy ldlm_processing_policy_table[] = {
85         [LDLM_PLAIN] ldlm_process_plain_lock,
86         [LDLM_EXTENT] ldlm_process_extent_lock,
87 #ifdef __KERNEL__
88         [LDLM_FLOCK] ldlm_process_flock_lock,
89 #endif
90 };
91
92 ldlm_processing_policy ldlm_get_processing_policy(struct ldlm_resource *res)
93 {
94         return ldlm_processing_policy_table[res->lr_type];
95 }
96
97 void ldlm_register_intent(struct ldlm_namespace *ns, ldlm_res_policy arg)
98 {
99         ns->ns_policy = arg;
100 }
101
102 /*
103  * REFCOUNTED LOCK OBJECTS
104  */
105
106
107 /*
108  * Lock refcounts, during creation:
109  *   - one special one for allocation, dec'd only once in destroy
110  *   - one for being a lock that's in-use
111  *   - one for the addref associated with a new lock
112  */
113 struct ldlm_lock *ldlm_lock_get(struct ldlm_lock *lock)
114 {
115         atomic_inc(&lock->l_refc);
116         return lock;
117 }
118
119 void ldlm_lock_put(struct ldlm_lock *lock)
120 {
121         ENTRY;
122
123         if (atomic_dec_and_test(&lock->l_refc)) {
124                 struct ldlm_namespace *ns = lock->l_resource->lr_namespace;
125
126                 l_lock(&ns->ns_lock);
127                 LDLM_DEBUG(lock, "final lock_put on destroyed lock, freeing");
128                 LASSERT(lock->l_destroyed);
129                 LASSERT(list_empty(&lock->l_res_link));
130
131                 spin_lock(&ns->ns_counter_lock);
132                 ns->ns_locks--;
133                 spin_unlock(&ns->ns_counter_lock);
134
135                 ldlm_resource_putref(lock->l_resource);
136                 lock->l_resource = NULL;
137                 if (lock->l_export)
138                         class_export_put(lock->l_export);
139
140                 if (lock->l_parent)
141                         LDLM_LOCK_PUT(lock->l_parent);
142
143                 if (lock->l_lvb_data != NULL)
144                         OBD_FREE(lock->l_lvb_data, lock->l_lvb_len);
145
146                 OBD_SLAB_FREE(lock, ldlm_lock_slab, sizeof(*lock));
147                 l_unlock(&ns->ns_lock);
148         }
149
150         EXIT;
151 }
152
153 void ldlm_lock_remove_from_lru(struct ldlm_lock *lock)
154 {
155         ENTRY;
156         l_lock(&lock->l_resource->lr_namespace->ns_lock);
157         if (!list_empty(&lock->l_lru)) {
158                 list_del_init(&lock->l_lru);
159                 lock->l_resource->lr_namespace->ns_nr_unused--;
160                 LASSERT(lock->l_resource->lr_namespace->ns_nr_unused >= 0);
161         }
162         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
163         EXIT;
164 }
165
166 /* This used to have a 'strict' flact, which recovery would use to mark an
167  * in-use lock as needing-to-die.  Lest I am ever tempted to put it back, I
168  * shall explain why it's gone: with the new hash table scheme, once you call
169  * ldlm_lock_destroy, you can never drop your final references on this lock.
170  * Because it's not in the hash table anymore.  -phil */
171 void ldlm_lock_destroy(struct ldlm_lock *lock)
172 {
173         ENTRY;
174         l_lock(&lock->l_resource->lr_namespace->ns_lock);
175
176         if (!list_empty(&lock->l_children)) {
177                 LDLM_ERROR(lock, "still has children (%p)!",
178                            lock->l_children.next);
179                 ldlm_lock_dump(D_ERROR, lock, 0);
180                 LBUG();
181         }
182         if (lock->l_readers || lock->l_writers) {
183                 LDLM_ERROR(lock, "lock still has references");
184                 ldlm_lock_dump(D_ERROR, lock, 0);
185                 LBUG();
186         }
187
188         if (!list_empty(&lock->l_res_link)) {
189                 ldlm_lock_dump(D_ERROR, lock, 0);
190                 LBUG();
191         }
192
193         if (lock->l_destroyed) {
194                 LASSERT(list_empty(&lock->l_lru));
195                 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
196                 EXIT;
197                 return;
198         }
199         lock->l_destroyed = 1;
200
201         list_del_init(&lock->l_export_chain);
202         ldlm_lock_remove_from_lru(lock);
203         class_handle_unhash(&lock->l_handle);
204
205 #if 0
206         /* Wake anyone waiting for this lock */
207         /* FIXME: I should probably add yet another flag, instead of using
208          * l_export to only call this on clients */
209         if (lock->l_export)
210                 class_export_put(lock->l_export);
211         lock->l_export = NULL;
212         if (lock->l_export && lock->l_completion_ast)
213                 lock->l_completion_ast(lock, 0);
214 #endif
215
216         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
217         LDLM_LOCK_PUT(lock);
218         EXIT;
219 }
220
221 /* this is called by portals_handle2object with the handle lock taken */
222 static void lock_handle_addref(void *lock)
223 {
224         LDLM_LOCK_GET((struct ldlm_lock *)lock);
225 }
226
227 /*
228  * usage: pass in a resource on which you have done ldlm_resource_get
229  *        pass in a parent lock on which you have done a ldlm_lock_get
230  *        after return, ldlm_*_put the resource and parent
231  * returns: lock with refcount 1
232  */
233 static struct ldlm_lock *ldlm_lock_new(struct ldlm_lock *parent,
234                                        struct ldlm_resource *resource)
235 {
236         struct ldlm_lock *lock;
237         ENTRY;
238
239         if (resource == NULL)
240                 LBUG();
241
242         OBD_SLAB_ALLOC(lock, ldlm_lock_slab, SLAB_NOFS, sizeof(*lock));
243         if (lock == NULL)
244                 RETURN(NULL);
245
246         lock->l_resource = ldlm_resource_getref(resource);
247
248         atomic_set(&lock->l_refc, 2);
249         INIT_LIST_HEAD(&lock->l_children);
250         INIT_LIST_HEAD(&lock->l_res_link);
251         INIT_LIST_HEAD(&lock->l_lru);
252         INIT_LIST_HEAD(&lock->l_export_chain);
253         INIT_LIST_HEAD(&lock->l_pending_chain);
254         init_waitqueue_head(&lock->l_waitq);
255
256         spin_lock(&resource->lr_namespace->ns_counter_lock);
257         resource->lr_namespace->ns_locks++;
258         spin_unlock(&resource->lr_namespace->ns_counter_lock);
259
260         if (parent != NULL) {
261                 l_lock(&parent->l_resource->lr_namespace->ns_lock);
262                 lock->l_parent = LDLM_LOCK_GET(parent);
263                 list_add(&lock->l_childof, &parent->l_children);
264                 l_unlock(&parent->l_resource->lr_namespace->ns_lock);
265         }
266
267         INIT_LIST_HEAD(&lock->l_handle.h_link);
268         class_handle_hash(&lock->l_handle, lock_handle_addref);
269
270         RETURN(lock);
271 }
272
273 int ldlm_lock_change_resource(struct ldlm_namespace *ns, struct ldlm_lock *lock,
274                               struct ldlm_res_id new_resid)
275 {
276         struct ldlm_resource *oldres = lock->l_resource;
277         ENTRY;
278
279         l_lock(&ns->ns_lock);
280         if (memcmp(&new_resid, &lock->l_resource->lr_name,
281                    sizeof(lock->l_resource->lr_name)) == 0) {
282                 /* Nothing to do */
283                 l_unlock(&ns->ns_lock);
284                 RETURN(0);
285         }
286
287         LASSERT(new_resid.name[0] != 0);
288
289         /* This function assumes that the lock isn't on any lists */
290         LASSERT(list_empty(&lock->l_res_link));
291
292         lock->l_resource = ldlm_resource_get(ns, NULL, new_resid,
293                                              lock->l_resource->lr_type, 1);
294         if (lock->l_resource == NULL) {
295                 LBUG();
296                 RETURN(-ENOMEM);
297         }
298
299         /* ...and the flowers are still standing! */
300         ldlm_resource_putref(oldres);
301
302         l_unlock(&ns->ns_lock);
303         RETURN(0);
304 }
305
306 /*
307  *  HANDLES
308  */
309
310 void ldlm_lock2handle(struct ldlm_lock *lock, struct lustre_handle *lockh)
311 {
312         lockh->cookie = lock->l_handle.h_cookie;
313 }
314
315 /* if flags: atomically get the lock and set the flags.
316  *           Return NULL if flag already set
317  */
318
319 struct ldlm_lock *__ldlm_handle2lock(struct lustre_handle *handle, int flags)
320 {
321         struct ldlm_namespace *ns;
322         struct ldlm_lock *lock = NULL, *retval = NULL;
323         ENTRY;
324
325         LASSERT(handle);
326
327         lock = class_handle2object(handle->cookie);
328         if (lock == NULL)
329                 RETURN(NULL);
330
331         LASSERT(lock->l_resource != NULL);
332         ns = lock->l_resource->lr_namespace;
333         LASSERT(ns != NULL);
334
335         l_lock(&ns->ns_lock);
336
337         /* It's unlikely but possible that someone marked the lock as
338          * destroyed after we did handle2object on it */
339         if (lock->l_destroyed) {
340                 CDEBUG(D_INFO, "lock already destroyed: lock %p\n", lock);
341                 LDLM_LOCK_PUT(lock);
342                 GOTO(out, retval);
343         }
344
345         if (flags && (lock->l_flags & flags)) {
346                 LDLM_LOCK_PUT(lock);
347                 GOTO(out, retval);
348         }
349
350         if (flags)
351                 lock->l_flags |= flags;
352
353         retval = lock;
354         EXIT;
355  out:
356         l_unlock(&ns->ns_lock);
357         return retval;
358 }
359
360 struct ldlm_lock *ldlm_handle2lock_ns(struct ldlm_namespace *ns,
361                                       struct lustre_handle *handle)
362 {
363         struct ldlm_lock *retval = NULL;
364
365         l_lock(&ns->ns_lock);
366         retval = __ldlm_handle2lock(handle, 0);
367         l_unlock(&ns->ns_lock);
368
369         return retval;
370 }
371
372 void ldlm_lock2desc(struct ldlm_lock *lock, struct ldlm_lock_desc *desc)
373 {
374         ldlm_res2desc(lock->l_resource, &desc->l_resource);
375         desc->l_req_mode = lock->l_req_mode;
376         desc->l_granted_mode = lock->l_granted_mode;
377         memcpy(&desc->l_policy_data, &lock->l_policy_data,
378                sizeof(desc->l_policy_data));
379 }
380
381 void ldlm_add_ast_work_item(struct ldlm_lock *lock, struct ldlm_lock *new,
382                             void *data, int datalen)
383 {
384         struct ldlm_ast_work *w;
385         ENTRY;
386
387         l_lock(&lock->l_resource->lr_namespace->ns_lock);
388         if (new && (lock->l_flags & LDLM_FL_AST_SENT))
389                 GOTO(out, 0);
390
391         CDEBUG(D_OTHER, "lock %p incompatible; sending blocking AST.\n", lock);
392
393         OBD_ALLOC(w, sizeof(*w));
394         if (!w) {
395                 LBUG();
396                 GOTO(out, 0);
397         }
398
399         w->w_data = data;
400         w->w_datalen = datalen;
401         if (new) {
402                 LDLM_DEBUG(lock, "lock incompatible; sending blocking AST.");
403                 lock->l_flags |= LDLM_FL_AST_SENT;
404                 /* If the enqueuing client said so, tell the AST recipient to
405                  * discard dirty data, rather than writing back. */
406                 if (new->l_flags & LDLM_AST_DISCARD_DATA)
407                         lock->l_flags |= LDLM_FL_DISCARD_DATA;
408                 w->w_blocking = 1;
409                 ldlm_lock2desc(new, &w->w_desc);
410         }
411
412         w->w_lock = LDLM_LOCK_GET(lock);
413         list_add(&w->w_list, lock->l_resource->lr_tmp);
414         EXIT;
415  out:
416         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
417 }
418
419 void ldlm_lock_addref(struct lustre_handle *lockh, __u32 mode)
420 {
421         struct ldlm_lock *lock;
422
423         lock = ldlm_handle2lock(lockh);
424         ldlm_lock_addref_internal(lock, mode);
425         LDLM_LOCK_PUT(lock);
426 }
427
428 /* only called for local locks */
429 void ldlm_lock_addref_internal(struct ldlm_lock *lock, __u32 mode)
430 {
431         l_lock(&lock->l_resource->lr_namespace->ns_lock);
432         ldlm_lock_remove_from_lru(lock);
433         if (mode & (LCK_NL | LCK_CR | LCK_PR))
434                 lock->l_readers++;
435         if (mode & (LCK_EX | LCK_CW | LCK_PW))
436                 lock->l_writers++;
437         lock->l_last_used = jiffies;
438         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
439         LDLM_LOCK_GET(lock);
440         LDLM_DEBUG(lock, "ldlm_lock_addref(%s)", ldlm_lockname[mode]);
441 }
442
443 void ldlm_lock_decref_internal(struct ldlm_lock *lock, __u32 mode)
444 {
445         struct ldlm_namespace *ns;
446         ENTRY;
447
448         LDLM_DEBUG(lock, "ldlm_lock_decref(%s)", ldlm_lockname[mode]);
449         ns = lock->l_resource->lr_namespace;
450         l_lock(&ns->ns_lock);
451         if (mode & (LCK_NL | LCK_CR | LCK_PR)) {
452                 LASSERT(lock->l_readers > 0);
453                 lock->l_readers--;
454         }
455         if (mode & (LCK_EX | LCK_CW | LCK_PW)) {
456                 LASSERT(lock->l_writers > 0);
457                 lock->l_writers--;
458         }
459
460         if (lock->l_flags & LDLM_FL_LOCAL &&
461             !lock->l_readers && !lock->l_writers) {
462                 /* If this is a local lock on a server namespace and this was
463                  * the last reference, cancel the lock. */
464                 CDEBUG(D_INFO, "forcing cancel of local lock\n");
465                 lock->l_flags |= LDLM_FL_CBPENDING;
466         }
467
468         if (!lock->l_readers && !lock->l_writers &&
469             (lock->l_flags & LDLM_FL_CBPENDING)) {
470                 /* If we received a blocked AST and this was the last reference,
471                  * run the callback. */
472                 if (ns->ns_client == LDLM_NAMESPACE_SERVER && lock->l_export)
473                         CERROR("FL_CBPENDING set on non-local lock--just a "
474                                "warning\n");
475
476                 LDLM_DEBUG(lock, "final decref done on cbpending lock");
477
478                 LDLM_LOCK_GET(lock); /* dropped by bl thread */
479                 ldlm_lock_remove_from_lru(lock);
480                 ldlm_bl_to_thread(ns, NULL, lock);
481                 l_unlock(&ns->ns_lock);
482         } else if (ns->ns_client == LDLM_NAMESPACE_CLIENT &&
483                    !lock->l_readers && !lock->l_writers) {
484                 /* If this is a client-side namespace and this was the last
485                  * reference, put it on the LRU. */
486                 LASSERT(list_empty(&lock->l_lru));
487                 LASSERT(ns->ns_nr_unused >= 0);
488                 list_add_tail(&lock->l_lru, &ns->ns_unused_list);
489                 ns->ns_nr_unused++;
490                 l_unlock(&ns->ns_lock);
491                 ldlm_cancel_lru(ns);
492         } else {
493                 l_unlock(&ns->ns_lock);
494         }
495
496         LDLM_LOCK_PUT(lock);    /* matches the ldlm_lock_get in addref */
497
498         EXIT;
499 }
500
501 void ldlm_lock_decref(struct lustre_handle *lockh, __u32 mode)
502 {
503         struct ldlm_lock *lock = __ldlm_handle2lock(lockh, 0);
504         LASSERT(lock != NULL);
505         ldlm_lock_decref_internal(lock, mode);
506         LDLM_LOCK_PUT(lock);
507 }
508
509 /* This will drop a lock reference and mark it for destruction, but will not
510  * necessarily cancel the lock before returning. */
511 void ldlm_lock_decref_and_cancel(struct lustre_handle *lockh, __u32 mode)
512 {
513         struct ldlm_lock *lock = __ldlm_handle2lock(lockh, 0);
514         ENTRY;
515
516         LASSERT(lock != NULL);
517
518         LDLM_DEBUG(lock, "ldlm_lock_decref(%s)", ldlm_lockname[mode]);
519         l_lock(&lock->l_resource->lr_namespace->ns_lock);
520         lock->l_flags |= LDLM_FL_CBPENDING;
521         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
522         ldlm_lock_decref_internal(lock, mode);
523         LDLM_LOCK_PUT(lock);
524 }
525
526 /* NOTE: called by
527  *  - ldlm_lock_enqueue
528  *  - ldlm_reprocess_queue
529  *  - ldlm_lock_convert
530  */
531 void ldlm_grant_lock(struct ldlm_lock *lock, void *data, int datalen,
532                      int run_ast)
533 {
534         struct ldlm_resource *res = lock->l_resource;
535         ENTRY;
536
537         l_lock(&lock->l_resource->lr_namespace->ns_lock);
538         lock->l_granted_mode = lock->l_req_mode;
539         ldlm_resource_add_lock(res, &res->lr_granted, lock);
540
541         if (lock->l_granted_mode < res->lr_most_restr)
542                 res->lr_most_restr = lock->l_granted_mode;
543
544         if (run_ast && lock->l_completion_ast != NULL)
545                 ldlm_add_ast_work_item(lock, NULL, data, datalen);
546
547         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
548         EXIT;
549 }
550
551 /* returns a referenced lock or NULL.  See the flag descriptions below, in the
552  * comment above ldlm_lock_match */
553 static struct ldlm_lock *search_queue(struct list_head *queue, ldlm_mode_t mode,
554                                       ldlm_policy_data_t *policy,
555                                       struct ldlm_lock *old_lock, int flags)
556 {
557         struct ldlm_lock *lock;
558         struct list_head *tmp;
559
560         list_for_each(tmp, queue) {
561                 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
562
563                 if (lock == old_lock)
564                         break;
565
566                 /* llite sometimes wants to match locks that will be
567                  * canceled when their users drop, but we allow it to match
568                  * if it passes in CBPENDING and the lock still has users.
569                  * this is generally only going to be used by children 
570                  * whose parents already hold a lock so forward progress
571                  * can still happen. */
572                 if (lock->l_flags & LDLM_FL_CBPENDING &&
573                     !(flags & LDLM_FL_CBPENDING))
574                         continue;
575                 if (lock->l_flags & LDLM_FL_CBPENDING &&
576                     lock->l_readers == 0 && lock->l_writers == 0)
577                         continue;
578
579                 if (!(lock->l_req_mode & mode))
580                         continue;
581
582                 if (lock->l_resource->lr_type == LDLM_EXTENT &&
583                     (lock->l_policy_data.l_extent.start >
584                      policy->l_extent.start ||
585                      lock->l_policy_data.l_extent.end < policy->l_extent.end))
586                         continue;
587
588                 if (lock->l_destroyed)
589                         continue;
590
591                 if ((flags & LDLM_FL_LOCAL_ONLY) &&
592                     !(lock->l_flags & LDLM_FL_LOCAL))
593                         continue;
594
595                 if (flags & LDLM_FL_TEST_LOCK)
596                         LDLM_LOCK_GET(lock);
597                 else
598                         ldlm_lock_addref_internal(lock, mode);
599                 return lock;
600         }
601
602         return NULL;
603 }
604
605 void ldlm_lock_allow_match(struct ldlm_lock *lock)
606 {
607         l_lock(&lock->l_resource->lr_namespace->ns_lock);
608         lock->l_flags |= LDLM_FL_CAN_MATCH;
609         wake_up(&lock->l_waitq);
610         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
611 }
612
613 /* Can be called in two ways:
614  *
615  * If 'ns' is NULL, then lockh describes an existing lock that we want to look
616  * for a duplicate of.
617  *
618  * Otherwise, all of the fields must be filled in, to match against.
619  *
620  * If 'flags' contains LDLM_FL_LOCAL_ONLY, then only match local locks on the
621  *     server (ie, connh is NULL)
622  * If 'flags' contains LDLM_FL_BLOCK_GRANTED, then only locks on the granted
623  *     list will be considered
624  * If 'flags' contains LDLM_FL_CBPENDING, then locks that have been marked
625  *     to be canceled can still be matched as long as they still have reader
626  *     or writer refernces
627  * If 'flags' contains LDLM_FL_TEST_LOCK, then don't actually reference a lock,
628  *     just tell us if we would have matched.
629  *
630  * Returns 1 if it finds an already-existing lock that is compatible; in this
631  * case, lockh is filled in with a addref()ed lock
632  */
633 int ldlm_lock_match(struct ldlm_namespace *ns, int flags,
634                     struct ldlm_res_id *res_id, __u32 type,
635                     ldlm_policy_data_t *policy, ldlm_mode_t mode,
636                     struct lustre_handle *lockh)
637 {
638         struct ldlm_resource *res;
639         struct ldlm_lock *lock, *old_lock = NULL;
640         int rc = 0;
641         ENTRY;
642
643         if (ns == NULL) {
644                 old_lock = ldlm_handle2lock(lockh);
645                 LASSERT(old_lock);
646
647                 ns = old_lock->l_resource->lr_namespace;
648                 res_id = &old_lock->l_resource->lr_name;
649                 type = old_lock->l_resource->lr_type;
650                 mode = old_lock->l_req_mode;
651         }
652
653         res = ldlm_resource_get(ns, NULL, *res_id, type, 0);
654         if (res == NULL) {
655                 LASSERT(old_lock == NULL);
656                 RETURN(0);
657         }
658
659         l_lock(&ns->ns_lock);
660
661         lock = search_queue(&res->lr_granted, mode, policy, old_lock, flags);
662         if (lock != NULL)
663                 GOTO(out, rc = 1);
664         if (flags & LDLM_FL_BLOCK_GRANTED)
665                 GOTO(out, rc = 0);
666         lock = search_queue(&res->lr_converting, mode, policy, old_lock, flags);
667         if (lock != NULL)
668                 GOTO(out, rc = 1);
669         lock = search_queue(&res->lr_waiting, mode, policy, old_lock, flags);
670         if (lock != NULL)
671                 GOTO(out, rc = 1);
672
673         EXIT;
674  out:
675         ldlm_resource_putref(res);
676         l_unlock(&ns->ns_lock);
677
678         if (lock) {
679                 ldlm_lock2handle(lock, lockh);
680                 if (!(lock->l_flags & LDLM_FL_CAN_MATCH)) {
681                         struct l_wait_info lwi;
682                         if (lock->l_completion_ast)
683                                 lock->l_completion_ast(lock,
684                                                        LDLM_FL_WAIT_NOREPROC,
685                                                        NULL);
686
687                         lwi = LWI_TIMEOUT_INTR(obd_timeout*HZ, NULL,NULL,NULL);
688
689                         /* XXX FIXME see comment on CAN_MATCH in lustre_dlm.h */
690                         l_wait_event(lock->l_waitq,
691                                      (lock->l_flags & LDLM_FL_CAN_MATCH), &lwi);
692                 }
693         }
694         if (rc)
695                 LDLM_DEBUG(lock, "matched ("LPU64" "LPU64")",
696                            type == LDLM_PLAIN ? res_id->name[2] :
697                                 policy->l_extent.start,
698                            type == LDLM_PLAIN ? res_id->name[3] :
699                                 policy->l_extent.end);
700         else if (!(flags & LDLM_FL_TEST_LOCK)) /* less verbose for test-only */
701                 LDLM_DEBUG_NOLOCK("not matched ns %p type %u mode %u res "
702                                   LPU64"/"LPU64" ("LPU64" "LPU64")", ns,
703                                   type, mode, res_id->name[0], res_id->name[1],
704                                   type == LDLM_PLAIN ? res_id->name[2] :
705                                         policy->l_extent.start,
706                                   type == LDLM_PLAIN ? res_id->name[3] :
707                                         policy->l_extent.end);
708
709         if (old_lock)
710                 LDLM_LOCK_PUT(old_lock);
711         if (flags & LDLM_FL_TEST_LOCK && rc)
712                 LDLM_LOCK_PUT(lock);
713
714         return rc;
715 }
716
717 /* Returns a referenced lock */
718 struct ldlm_lock *ldlm_lock_create(struct ldlm_namespace *ns,
719                                    struct lustre_handle *parent_lock_handle,
720                                    struct ldlm_res_id res_id, __u32 type,
721                                    ldlm_mode_t mode,
722                                    ldlm_blocking_callback blocking,
723                                    ldlm_completion_callback completion,
724                                    ldlm_glimpse_callback glimpse,
725                                    void *data, __u32 lvb_len)
726 {
727         struct ldlm_resource *res, *parent_res = NULL;
728         struct ldlm_lock *lock, *parent_lock = NULL;
729         ENTRY;
730
731         if (parent_lock_handle) {
732                 parent_lock = ldlm_handle2lock(parent_lock_handle);
733                 if (parent_lock)
734                         parent_res = parent_lock->l_resource;
735         }
736
737         res = ldlm_resource_get(ns, parent_res, res_id, type, 1);
738         if (res == NULL)
739                 RETURN(NULL);
740
741         lock = ldlm_lock_new(parent_lock, res);
742         ldlm_resource_putref(res);
743         if (parent_lock != NULL)
744                 LDLM_LOCK_PUT(parent_lock);
745
746         if (lock == NULL)
747                 RETURN(NULL);
748
749         lock->l_req_mode = mode;
750         lock->l_ast_data = data;
751         lock->l_blocking_ast = blocking;
752         lock->l_completion_ast = completion;
753         lock->l_glimpse_ast = glimpse;
754
755         lock->l_lvb_len = lvb_len;
756         OBD_ALLOC(lock->l_lvb_data, lvb_len);
757         if (lock->l_lvb_data == NULL) {
758                 OBD_SLAB_FREE(lock, ldlm_lock_slab, sizeof(*lock));
759                 RETURN(NULL);
760         }
761
762         RETURN(lock);
763 }
764
765 ldlm_error_t ldlm_lock_enqueue(struct ldlm_namespace *ns,
766                                struct ldlm_lock **lockp,
767                                void *cookie, int *flags)
768 {
769         struct ldlm_lock *lock = *lockp;
770         struct ldlm_resource *res = lock->l_resource;
771         int local = res->lr_namespace->ns_client;
772         ldlm_processing_policy policy;
773         ldlm_error_t rc = ELDLM_OK;
774         ENTRY;
775
776         /* policies are not executed on the client or during replay */
777         if ((*flags & (LDLM_FL_HAS_INTENT|LDLM_FL_REPLAY)) == LDLM_FL_HAS_INTENT
778             && !local && ns->ns_policy) {
779                 rc = ns->ns_policy(ns, lockp, cookie, lock->l_req_mode, *flags,
780                                    NULL);
781                 if (rc == ELDLM_LOCK_REPLACED) {
782                         /* The lock that was returned has already been granted,
783                          * and placed into lockp.  If it's not the same as the
784                          * one we passed in, then destroy the old one and our
785                          * work here is done. */
786                         if (lock != *lockp) {
787                                 ldlm_lock_destroy(lock);
788                                 LDLM_LOCK_PUT(lock);
789                         }
790                         *flags |= LDLM_FL_LOCK_CHANGED;
791                         RETURN(0);
792                 } else if (rc == ELDLM_LOCK_ABORTED ||
793                            (rc == 0 && (*flags & LDLM_FL_INTENT_ONLY))) {
794                         ldlm_lock_destroy(lock);
795                         RETURN(rc);
796                 }
797                 LASSERT(rc == ELDLM_OK);
798         }
799
800         l_lock(&ns->ns_lock);
801         if (local && lock->l_req_mode == lock->l_granted_mode) {
802                 /* The server returned a blocked lock, but it was granted before
803                  * we got a chance to actually enqueue it.  We don't need to do
804                  * anything else. */
805                 *flags &= ~(LDLM_FL_BLOCK_GRANTED |
806                             LDLM_FL_BLOCK_CONV | LDLM_FL_BLOCK_WAIT);
807                 GOTO(out, ELDLM_OK);
808         }
809
810         /* Some flags from the enqueue want to make it into the AST, via the
811          * lock's l_flags. */
812         lock->l_flags |= (*flags & LDLM_AST_DISCARD_DATA);
813
814         /* This distinction between local lock trees is very important; a client
815          * namespace only has information about locks taken by that client, and
816          * thus doesn't have enough information to decide for itself if it can
817          * be granted (below).  In this case, we do exactly what the server
818          * tells us to do, as dictated by the 'flags'.
819          *
820          * We do exactly the same thing during recovery, when the server is
821          * more or less trusting the clients not to lie.
822          *
823          * FIXME (bug 268): Detect obvious lies by checking compatibility in
824          * granted/converting queues. */
825         ldlm_resource_unlink_lock(lock);
826         if (local) {
827                 if (*flags & LDLM_FL_BLOCK_CONV)
828                         ldlm_resource_add_lock(res, &res->lr_converting, lock);
829                 else if (*flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED))
830                         ldlm_resource_add_lock(res, &res->lr_waiting, lock);
831                 else
832                         ldlm_grant_lock(lock, NULL, 0, 0);
833                 GOTO(out, ELDLM_OK);
834         } else if (*flags & LDLM_FL_REPLAY) {
835                 if (*flags & LDLM_FL_BLOCK_CONV) {
836                         ldlm_resource_add_lock(res, &res->lr_converting, lock);
837                         GOTO(out, ELDLM_OK);
838                 } else if (*flags & LDLM_FL_BLOCK_WAIT) {
839                         ldlm_resource_add_lock(res, &res->lr_waiting, lock);
840                         GOTO(out, ELDLM_OK);
841                 } else if (*flags & LDLM_FL_BLOCK_GRANTED) {
842                         ldlm_grant_lock(lock, NULL, 0, 0);
843                         GOTO(out, ELDLM_OK);
844                 }
845                 /* If no flags, fall through to normal enqueue path. */
846         }
847
848         policy = ldlm_processing_policy_table[res->lr_type];
849         policy(lock, flags, 1, &rc);
850         EXIT;
851 out:
852         l_unlock(&ns->ns_lock);
853         return rc;
854 }
855
856 /* Must be called with namespace taken: queue is waiting or converting. */
857 int ldlm_reprocess_queue(struct ldlm_resource *res, struct list_head *queue)
858 {
859         struct list_head *tmp, *pos;
860         ldlm_processing_policy policy;
861         int flags;
862         int rc = LDLM_ITER_CONTINUE;
863         ldlm_error_t err;
864         ENTRY;
865
866         policy = ldlm_processing_policy_table[res->lr_type];
867         LASSERT(policy);
868
869         list_for_each_safe(tmp, pos, queue) {
870                 struct ldlm_lock *pending;
871                 pending = list_entry(tmp, struct ldlm_lock, l_res_link);
872
873                 CDEBUG(D_INFO, "Reprocessing lock %p\n", pending);
874
875                 flags = 0;
876                 rc = policy(pending, &flags, 0, &err);
877                 if (rc != LDLM_ITER_CONTINUE)
878                         break;
879         }
880
881         RETURN(rc);
882 }
883
884 int ldlm_run_ast_work(struct ldlm_namespace *ns, struct list_head *rpc_list)
885 {
886         struct list_head *tmp, *pos;
887         int rc, retval = 0;
888         ENTRY;
889
890         l_check_no_ns_lock(ns);
891
892         list_for_each_safe(tmp, pos, rpc_list) {
893                 struct ldlm_ast_work *w =
894                         list_entry(tmp, struct ldlm_ast_work, w_list);
895
896                 /* It's possible to receive a completion AST before we've set
897                  * the l_completion_ast pointer: either because the AST arrived
898                  * before the reply, or simply because there's a small race
899                  * window between receiving the reply and finishing the local
900                  * enqueue. (bug 842)
901                  *
902                  * This can't happen with the blocking_ast, however, because we
903                  * will never call the local blocking_ast until we drop our
904                  * reader/writer reference, which we won't do until we get the
905                  * reply and finish enqueueing. */
906                 LASSERT(w->w_lock != NULL);
907                 if (w->w_blocking) {
908                         LASSERT(w->w_lock->l_blocking_ast != NULL);
909                         rc = w->w_lock->l_blocking_ast
910                                 (w->w_lock, &w->w_desc, w->w_data,
911                                  LDLM_CB_BLOCKING);
912                 } else if (w->w_lock->l_completion_ast != NULL) {
913                         LASSERT(w->w_lock->l_completion_ast != NULL);
914                         rc = w->w_lock->l_completion_ast(w->w_lock, w->w_flags,
915                                                          w->w_data);
916                 } else {
917                         rc = 0;
918                 }
919                 if (rc == -ERESTART)
920                         retval = rc;
921                 else if (rc)
922                         CDEBUG(D_DLMTRACE, "Failed AST - should clean & "
923                                "disconnect client\n");
924                 LDLM_LOCK_PUT(w->w_lock);
925                 list_del(&w->w_list);
926                 OBD_FREE(w, sizeof(*w));
927         }
928         RETURN(retval);
929 }
930
931 static int reprocess_one_queue(struct ldlm_resource *res, void *closure)
932 {
933         ldlm_reprocess_all(res);
934         return LDLM_ITER_CONTINUE;
935 }
936
937 void ldlm_reprocess_all_ns(struct ldlm_namespace *ns)
938 {
939         int i, rc;
940
941         l_lock(&ns->ns_lock);
942         for (i = 0; i < RES_HASH_SIZE; i++) {
943                 struct list_head *tmp, *next;
944                 list_for_each_safe(tmp, next, &(ns->ns_hash[i])) {
945                         struct ldlm_resource *res =
946                                 list_entry(tmp, struct ldlm_resource, lr_hash);
947
948                         ldlm_resource_getref(res);
949                         l_unlock(&ns->ns_lock);
950                         rc = reprocess_one_queue(res, NULL);
951                         l_lock(&ns->ns_lock);
952                         next = tmp->next;
953                         ldlm_resource_putref(res);
954                         if (rc == LDLM_ITER_STOP)
955                                 GOTO(out, rc);
956                 }
957         }
958  out:
959         l_unlock(&ns->ns_lock);
960         EXIT;
961 }
962
963 void ldlm_reprocess_all(struct ldlm_resource *res)
964 {
965         struct list_head rpc_list = LIST_HEAD_INIT(rpc_list);
966         int rc;
967         ENTRY;
968
969         /* Local lock trees don't get reprocessed. */
970         if (res->lr_namespace->ns_client) {
971                 EXIT;
972                 return;
973         }
974
975  restart:
976         l_lock(&res->lr_namespace->ns_lock);
977         res->lr_tmp = &rpc_list;
978
979         rc = ldlm_reprocess_queue(res, &res->lr_converting);
980         if (rc == LDLM_ITER_CONTINUE)
981                 ldlm_reprocess_queue(res, &res->lr_waiting);
982
983         res->lr_tmp = NULL;
984         l_unlock(&res->lr_namespace->ns_lock);
985
986         rc = ldlm_run_ast_work(res->lr_namespace, &rpc_list);
987         if (rc == -ERESTART) {
988                 LASSERT(list_empty(&rpc_list));
989                 goto restart;
990         }
991         EXIT;
992 }
993
994 void ldlm_cancel_callback(struct ldlm_lock *lock)
995 {
996         l_lock(&lock->l_resource->lr_namespace->ns_lock);
997         if (!(lock->l_flags & LDLM_FL_CANCEL)) {
998                 lock->l_flags |= LDLM_FL_CANCEL;
999                 if (lock->l_blocking_ast) {
1000                         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
1001                         // l_check_no_ns_lock(lock->l_resource->lr_namespace);
1002                         lock->l_blocking_ast(lock, NULL, lock->l_ast_data,
1003                                              LDLM_CB_CANCELING);
1004                         return;
1005                 } else {
1006                         LDLM_DEBUG(lock, "no blocking ast");
1007                 }
1008         }
1009         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
1010 }
1011
1012 void ldlm_lock_cancel(struct ldlm_lock *lock)
1013 {
1014         struct ldlm_resource *res;
1015         struct ldlm_namespace *ns;
1016         ENTRY;
1017
1018         /* There's no race between calling this and taking the ns lock below;
1019          * a lock can only be put on the waiting list once, because it can only
1020          * issue a blocking AST once. */
1021         ldlm_del_waiting_lock(lock);
1022
1023         res = lock->l_resource;
1024         ns = res->lr_namespace;
1025
1026         l_lock(&ns->ns_lock);
1027         /* Please do not, no matter how tempting, remove this LBUG without
1028          * talking to me first. -phik */
1029         if (lock->l_readers || lock->l_writers) {
1030                 LDLM_DEBUG(lock, "lock still has references");
1031                 ldlm_lock_dump(D_OTHER, lock, 0);
1032                 LBUG();
1033         }
1034
1035         ldlm_cancel_callback(lock);
1036
1037         ldlm_resource_unlink_lock(lock);
1038         ldlm_lock_destroy(lock);
1039         l_unlock(&ns->ns_lock);
1040         EXIT;
1041 }
1042
1043 int ldlm_lock_set_data(struct lustre_handle *lockh, void *data)
1044 {
1045         struct ldlm_lock *lock = ldlm_handle2lock(lockh);
1046         ENTRY;
1047
1048         if (lock == NULL)
1049                 RETURN(-EINVAL);
1050
1051         lock->l_ast_data = data;
1052         LDLM_LOCK_PUT(lock);
1053         RETURN(0);
1054 }
1055
1056 void ldlm_cancel_locks_for_export(struct obd_export *exp)
1057 {
1058         struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
1059         struct ldlm_lock *lock;
1060         struct ldlm_resource *res;
1061
1062         l_lock(&ns->ns_lock);
1063         while(!list_empty(&exp->exp_ldlm_data.led_held_locks)) { 
1064                 lock = list_entry(exp->exp_ldlm_data.led_held_locks.next,
1065                                   struct ldlm_lock, l_export_chain);
1066                 res = ldlm_resource_getref(lock->l_resource);
1067                 LDLM_DEBUG(lock, "export %p", exp);
1068                 ldlm_lock_cancel(lock);
1069                 l_unlock(&ns->ns_lock);
1070                 ldlm_reprocess_all(res);
1071                 ldlm_resource_putref(res);
1072                 l_lock(&ns->ns_lock);
1073         }
1074         l_unlock(&ns->ns_lock);
1075 }
1076
1077 struct ldlm_resource *ldlm_lock_convert(struct ldlm_lock *lock, int new_mode,
1078                                         int *flags)
1079 {
1080         struct list_head rpc_list = LIST_HEAD_INIT(rpc_list);
1081         struct ldlm_resource *res;
1082         struct ldlm_namespace *ns;
1083         int granted = 0;
1084         ENTRY;
1085
1086         LBUG();
1087
1088         res = lock->l_resource;
1089         ns = res->lr_namespace;
1090
1091         l_lock(&ns->ns_lock);
1092
1093         lock->l_req_mode = new_mode;
1094         ldlm_resource_unlink_lock(lock);
1095
1096         /* If this is a local resource, put it on the appropriate list. */
1097         if (res->lr_namespace->ns_client) {
1098                 if (*flags & (LDLM_FL_BLOCK_CONV | LDLM_FL_BLOCK_GRANTED)) {
1099                         ldlm_resource_add_lock(res, &res->lr_converting, lock);
1100                 } else {
1101                         /* This should never happen, because of the way the
1102                          * server handles conversions. */
1103                         LBUG();
1104
1105                         res->lr_tmp = &rpc_list;
1106                         ldlm_grant_lock(lock, NULL, 0, 0);
1107                         res->lr_tmp = NULL;
1108                         granted = 1;
1109                         /* FIXME: completion handling not with ns_lock held ! */
1110                         if (lock->l_completion_ast)
1111                                 lock->l_completion_ast(lock, 0, NULL);
1112                 }
1113         } else {
1114                 /* FIXME: We should try the conversion right away and possibly
1115                  * return success without the need for an extra AST */
1116                 ldlm_resource_add_lock(res, &res->lr_converting, lock);
1117                 *flags |= LDLM_FL_BLOCK_CONV;
1118         }
1119
1120         l_unlock(&ns->ns_lock);
1121
1122         if (granted)
1123                 ldlm_run_ast_work(ns, &rpc_list);
1124         RETURN(res);
1125 }
1126
1127 void ldlm_lock_dump(int level, struct ldlm_lock *lock, int pos)
1128 {
1129         char str[PTL_NALFMT_SIZE];
1130         struct obd_device *obd = NULL;
1131
1132         if (!((portal_debug | D_ERROR) & level))
1133                 return;
1134
1135         if (!lock) {
1136                 CDEBUG(level, "  NULL LDLM lock\n");
1137                 return;
1138         }
1139
1140         CDEBUG(level, "  -- Lock dump: %p/"LPX64" (rc: %d) (pos: %d)\n",
1141                lock, lock->l_handle.h_cookie, atomic_read(&lock->l_refc),
1142                pos);
1143         if (lock->l_conn_export != NULL)
1144                 obd = lock->l_conn_export->exp_obd;
1145         if (lock->l_export && lock->l_export->exp_connection) {
1146                 CDEBUG(level, "  Node: NID "LPX64" (%s) on %s (rhandle: "LPX64")\n",
1147                        lock->l_export->exp_connection->c_peer.peer_nid,
1148                        portals_nid2str(lock->l_export->exp_connection->c_peer.peer_ni->pni_number,
1149                                        lock->l_export->exp_connection->c_peer.peer_nid, str),
1150                        lock->l_export->exp_connection->c_peer.peer_ni->pni_name,
1151                        lock->l_remote_handle.cookie);
1152         } else if (obd == NULL) {
1153                 CDEBUG(level, "  Node: local\n");
1154         } else {
1155                 struct obd_import *imp = obd->u.cli.cl_import;
1156                 CDEBUG(level, "  Node: NID "LPX64" (%s) on %s (rhandle: "LPX64")\n",
1157                        imp->imp_connection->c_peer.peer_nid,
1158                        portals_nid2str(imp->imp_connection->c_peer.peer_ni->pni_number,
1159                                        imp->imp_connection->c_peer.peer_nid, str),
1160                        imp->imp_connection->c_peer.peer_ni->pni_name,
1161                        lock->l_remote_handle.cookie);
1162         }
1163         CDEBUG(level, "  Resource: %p ("LPU64"/"LPU64")\n", lock->l_resource,
1164                lock->l_resource->lr_name.name[0],
1165                lock->l_resource->lr_name.name[1]);
1166         CDEBUG(level, "  Req mode: %d, grant mode: %d, rc: %u, read: %d, "
1167                "write: %d\n", (int)lock->l_req_mode, (int)lock->l_granted_mode,
1168                atomic_read(&lock->l_refc), lock->l_readers, lock->l_writers);
1169         if (lock->l_resource->lr_type == LDLM_EXTENT)
1170                 CDEBUG(level, "  Extent: "LPU64" -> "LPU64
1171                        " (req "LPU64"-"LPU64")\n",
1172                        lock->l_policy_data.l_extent.start,
1173                        lock->l_policy_data.l_extent.end,
1174                        lock->l_req_extent.start, lock->l_req_extent.end);
1175         else if (lock->l_resource->lr_type == LDLM_FLOCK)
1176                 CDEBUG(level, "  Pid: %d Extent: "LPU64" -> "LPU64"\n",
1177                        lock->l_policy_data.l_flock.pid,
1178                        lock->l_policy_data.l_flock.start,
1179                        lock->l_policy_data.l_flock.end);
1180 }
1181
1182 void ldlm_lock_dump_handle(int level, struct lustre_handle *lockh)
1183 {
1184         struct ldlm_lock *lock;
1185
1186         lock = ldlm_handle2lock(lockh);
1187         if (lock == NULL)
1188                 return;
1189
1190         ldlm_lock_dump(D_OTHER, lock, 0);
1191
1192         LDLM_LOCK_PUT(lock);
1193 }