Whamcloud - gitweb
LU-12166 test: fix broken detection on ZFS
[fs/lustre-release.git] / lustre / ldlm / ldlm_flock.c
1 /*
2  * GPL HEADER START
3  *
4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
5  *
6  * This program is free software; you can redistribute it and/or modify
7  * it under the terms of the GNU General Public License version 2 only,
8  * as published by the Free Software Foundation.
9  *
10  * This program is distributed in the hope that it will be useful, but
11  * WITHOUT ANY WARRANTY; without even the implied warranty of
12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
13  * General Public License version 2 for more details (a copy is included
14  * in the LICENSE file that accompanied this code).
15  *
16  * You should have received a copy of the GNU General Public License
17  * version 2 along with this program; If not, see
18  * http://www.gnu.org/licenses/gpl-2.0.html
19  *
20  * GPL HEADER END
21  */
22 /*
23  * Copyright (c) 2003 Hewlett-Packard Development Company LP.
24  * Developed under the sponsorship of the US Government under
25  * Subcontract No. B514193
26  *
27  * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
28  * Use is subject to license terms.
29  *
30  * Copyright (c) 2010, 2017, Intel Corporation.
31  */
32 /*
33  * This file is part of Lustre, http://www.lustre.org/
34  * Lustre is a trademark of Sun Microsystems, Inc.
35  */
36
37 /**
38  * This file implements POSIX lock type for Lustre.
39  * Its policy properties are start and end of extent and PID.
40  *
41  * These locks are only done through MDS due to POSIX semantics requiring
42  * e.g. that locks could be only partially released and as such split into
43  * two parts, and also that two adjacent locks from the same process may be
44  * merged into a single wider lock.
45  *
46  * Lock modes are mapped like this:
47  * PR and PW for READ and WRITE locks
48  * NL to request a releasing of a portion of the lock
49  *
50  * These flock locks never timeout.
51  */
52
53 #define DEBUG_SUBSYSTEM S_LDLM
54
55 #include <linux/list.h>
56 #include <lustre_dlm.h>
57 #include <obd_support.h>
58 #include <obd_class.h>
59 #include <lustre_lib.h>
60
61 #include "ldlm_internal.h"
62
63 int ldlm_flock_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
64                             void *data, int flag);
65
66 /**
67  * list_for_remaining_safe - iterate over the remaining entries in a list
68  *              and safeguard against removal of a list entry.
69  * \param pos   the &struct list_head to use as a loop counter. pos MUST
70  *              have been initialized prior to using it in this macro.
71  * \param n     another &struct list_head to use as temporary storage
72  * \param head  the head for your list.
73  */
74 #define list_for_remaining_safe(pos, n, head) \
75         for (n = pos->next; pos != (head); pos = n, n = pos->next)
76
77 static inline int
78 ldlm_same_flock_owner(struct ldlm_lock *lock, struct ldlm_lock *new)
79 {
80         return((new->l_policy_data.l_flock.owner ==
81                 lock->l_policy_data.l_flock.owner) &&
82                (new->l_export == lock->l_export));
83 }
84
85 static inline int
86 ldlm_flocks_overlap(struct ldlm_lock *lock, struct ldlm_lock *new)
87 {
88         return((new->l_policy_data.l_flock.start <=
89                 lock->l_policy_data.l_flock.end) &&
90                (new->l_policy_data.l_flock.end >=
91                 lock->l_policy_data.l_flock.start));
92 }
93
94 static inline void ldlm_flock_blocking_link(struct ldlm_lock *req,
95                                             struct ldlm_lock *lock)
96 {
97         /* For server only */
98         if (req->l_export == NULL)
99                 return;
100
101         LASSERT(hlist_unhashed(&req->l_exp_flock_hash));
102
103         req->l_policy_data.l_flock.blocking_owner =
104                 lock->l_policy_data.l_flock.owner;
105         req->l_policy_data.l_flock.blocking_export =
106                 lock->l_export;
107         atomic_set(&req->l_policy_data.l_flock.blocking_refs, 0);
108
109         cfs_hash_add(req->l_export->exp_flock_hash,
110                      &req->l_policy_data.l_flock.owner,
111                      &req->l_exp_flock_hash);
112 }
113
114 static inline void ldlm_flock_blocking_unlink(struct ldlm_lock *req)
115 {
116         /* For server only */
117         if (req->l_export == NULL)
118                 return;
119
120         check_res_locked(req->l_resource);
121         if (req->l_export->exp_flock_hash != NULL &&
122             !hlist_unhashed(&req->l_exp_flock_hash))
123                 cfs_hash_del(req->l_export->exp_flock_hash,
124                              &req->l_policy_data.l_flock.owner,
125                              &req->l_exp_flock_hash);
126 }
127
128 static inline void
129 ldlm_flock_destroy(struct ldlm_lock *lock, enum ldlm_mode mode, __u64 flags)
130 {
131         ENTRY;
132
133         LDLM_DEBUG(lock, "ldlm_flock_destroy(mode: %d, flags: %#llx)",
134                    mode, flags);
135
136         /* Safe to not lock here, since it should be empty anyway */
137         LASSERT(hlist_unhashed(&lock->l_exp_flock_hash));
138
139         list_del_init(&lock->l_res_link);
140         if (flags == LDLM_FL_WAIT_NOREPROC) {
141                 /* client side - set a flag to prevent sending a CANCEL */
142                 lock->l_flags |= LDLM_FL_LOCAL_ONLY | LDLM_FL_CBPENDING;
143
144                 /* when reaching here, it is under lock_res_and_lock(). Thus,
145                    need call the nolock version of ldlm_lock_decref_internal*/
146                 ldlm_lock_decref_internal_nolock(lock, mode);
147         }
148
149         ldlm_lock_destroy_nolock(lock);
150         EXIT;
151 }
152
153 /**
154  * POSIX locks deadlock detection code.
155  *
156  * Given a new lock \a req and an existing lock \a bl_lock it conflicts
157  * with, we need to iterate through all blocked POSIX locks for this
158  * export and see if there is a deadlock condition arising. (i.e. when
159  * one client holds a lock on something and want a lock on something
160  * else and at the same time another client has the opposite situation).
161  */
162
163 struct ldlm_flock_lookup_cb_data {
164         __u64 *bl_owner;
165         struct ldlm_lock *lock;
166         struct obd_export *exp;
167 };
168
169 static int ldlm_flock_lookup_cb(struct cfs_hash *hs, struct cfs_hash_bd *bd,
170                                 struct hlist_node *hnode, void *data)
171 {
172         struct ldlm_flock_lookup_cb_data *cb_data = data;
173         struct obd_export *exp = cfs_hash_object(hs, hnode);
174         struct ldlm_lock *lock;
175
176         lock = cfs_hash_lookup(exp->exp_flock_hash, cb_data->bl_owner);
177         if (lock == NULL)
178                 return 0;
179
180         /* Stop on first found lock. Same process can't sleep twice */
181         cb_data->lock = lock;
182         cb_data->exp = class_export_get(exp);
183
184         return 1;
185 }
186
187 static int
188 ldlm_flock_deadlock(struct ldlm_lock *req, struct ldlm_lock *bl_lock)
189 {
190         struct obd_export *req_exp = req->l_export;
191         struct obd_export *bl_exp = bl_lock->l_export;
192         __u64 req_owner = req->l_policy_data.l_flock.owner;
193         __u64 bl_owner = bl_lock->l_policy_data.l_flock.owner;
194
195         /* For server only */
196         if (req_exp == NULL)
197                 return 0;
198
199         class_export_get(bl_exp);
200         while (1) {
201                 struct ldlm_flock_lookup_cb_data cb_data = {
202                                         .bl_owner = &bl_owner,
203                                         .lock = NULL,
204                                         .exp = NULL };
205                 struct obd_export *bl_exp_new;
206                 struct ldlm_lock *lock = NULL;
207                 struct ldlm_flock *flock;
208
209                 if (bl_exp->exp_flock_hash != NULL) {
210                         cfs_hash_for_each_key(bl_exp->exp_obd->obd_nid_hash,
211                                 &bl_exp->exp_connection->c_peer.nid,
212                                 ldlm_flock_lookup_cb, &cb_data);
213                         lock = cb_data.lock;
214                 }
215                 if (lock == NULL)
216                         break;
217
218                 class_export_put(bl_exp);
219                 bl_exp = cb_data.exp;
220
221                 LASSERT(req != lock);
222                 flock = &lock->l_policy_data.l_flock;
223                 LASSERT(flock->owner == bl_owner);
224                 bl_owner = flock->blocking_owner;
225                 bl_exp_new = class_export_get(flock->blocking_export);
226                 class_export_put(bl_exp);
227
228                 cfs_hash_put(bl_exp->exp_flock_hash, &lock->l_exp_flock_hash);
229                 bl_exp = bl_exp_new;
230
231                 if (bl_exp->exp_failed)
232                         break;
233
234                 if (bl_owner == req_owner &&
235                     (bl_exp->exp_connection->c_peer.nid ==
236                      req_exp->exp_connection->c_peer.nid)) {
237                         class_export_put(bl_exp);
238                         return 1;
239                 }
240         }
241         class_export_put(bl_exp);
242
243         return 0;
244 }
245
246 static void ldlm_flock_cancel_on_deadlock(struct ldlm_lock *lock,
247                                           struct list_head *work_list)
248 {
249         CDEBUG(D_INFO, "reprocess deadlock req=%p\n", lock);
250
251         if ((exp_connect_flags(lock->l_export) &
252                                 OBD_CONNECT_FLOCK_DEAD) == 0) {
253                 CERROR("deadlock found, but client doesn't "
254                                 "support flock canceliation\n");
255         } else {
256                 LASSERT(lock->l_completion_ast);
257                 LASSERT(!ldlm_is_ast_sent(lock));
258                 lock->l_flags |= LDLM_FL_AST_SENT | LDLM_FL_CANCEL_ON_BLOCK |
259                         LDLM_FL_FLOCK_DEADLOCK;
260                 ldlm_flock_blocking_unlink(lock);
261                 ldlm_resource_unlink_lock(lock);
262                 ldlm_add_ast_work_item(lock, NULL, work_list);
263         }
264 }
265
266 /**
267  * Process a granting attempt for flock lock.
268  * Must be called under ns lock held.
269  *
270  * This function looks for any conflicts for \a lock in the granted or
271  * waiting queues. The lock is granted if no conflicts are found in
272  * either queue.
273  */
274 int
275 ldlm_process_flock_lock(struct ldlm_lock *req, __u64 *flags,
276                         enum ldlm_process_intention intention,
277                         enum ldlm_error *err, struct list_head *work_list)
278 {
279         struct ldlm_resource *res = req->l_resource;
280         struct ldlm_namespace *ns = ldlm_res_to_ns(res);
281         struct list_head *tmp;
282         struct list_head *ownlocks = NULL;
283         struct ldlm_lock *lock = NULL;
284         struct ldlm_lock *new = req;
285         struct ldlm_lock *new2 = NULL;
286         enum ldlm_mode mode = req->l_req_mode;
287         int local = ns_is_client(ns);
288         int added = (mode == LCK_NL);
289         int overlaps = 0;
290         int splitted = 0;
291         const struct ldlm_callback_suite null_cbs = { NULL };
292         struct list_head *grant_work = intention == LDLM_PROCESS_ENQUEUE ?
293                                                         NULL : work_list;
294         ENTRY;
295
296         CDEBUG(D_DLMTRACE, "flags %#llx owner %llu pid %u mode %u start "
297                "%llu end %llu\n", *flags,
298                new->l_policy_data.l_flock.owner,
299                new->l_policy_data.l_flock.pid, mode,
300                req->l_policy_data.l_flock.start,
301                req->l_policy_data.l_flock.end);
302
303         *err = ELDLM_OK;
304
305         if (local) {
306                 /* No blocking ASTs are sent to the clients for
307                  * Posix file & record locks */
308                 req->l_blocking_ast = NULL;
309         } else {
310                 /* Called on the server for lock cancels. */
311                 req->l_blocking_ast = ldlm_flock_blocking_ast;
312         }
313
314 reprocess:
315         if ((*flags == LDLM_FL_WAIT_NOREPROC) || (mode == LCK_NL)) {
316                 /* This loop determines where this processes locks start
317                  * in the resource lr_granted list. */
318                 list_for_each(tmp, &res->lr_granted) {
319                         lock = list_entry(tmp, struct ldlm_lock,
320                                               l_res_link);
321                         if (ldlm_same_flock_owner(lock, req)) {
322                                 ownlocks = tmp;
323                                 break;
324                         }
325                 }
326         } else {
327                 int reprocess_failed = 0;
328                 lockmode_verify(mode);
329
330                 /* This loop determines if there are existing locks
331                  * that conflict with the new lock request. */
332                 list_for_each(tmp, &res->lr_granted) {
333                         lock = list_entry(tmp, struct ldlm_lock,
334                                               l_res_link);
335
336                         if (ldlm_same_flock_owner(lock, req)) {
337                                 if (!ownlocks)
338                                         ownlocks = tmp;
339                                 continue;
340                         }
341
342                         /* locks are compatible, overlap doesn't matter */
343                         if (lockmode_compat(lock->l_granted_mode, mode))
344                                 continue;
345
346                         if (!ldlm_flocks_overlap(lock, req))
347                                 continue;
348
349                         if (intention != LDLM_PROCESS_ENQUEUE) {
350                                 reprocess_failed = 1;
351                                 if (ldlm_flock_deadlock(req, lock)) {
352                                         ldlm_flock_cancel_on_deadlock(req,
353                                                         grant_work);
354                                         RETURN(LDLM_ITER_CONTINUE);
355                                 }
356                                 continue;
357                         }
358
359                         if (*flags & LDLM_FL_BLOCK_NOWAIT) {
360                                 ldlm_flock_destroy(req, mode, *flags);
361                                 *err = -EAGAIN;
362                                 RETURN(LDLM_ITER_STOP);
363                         }
364
365                         if (*flags & LDLM_FL_TEST_LOCK) {
366                                 ldlm_flock_destroy(req, mode, *flags);
367                                 req->l_req_mode = lock->l_granted_mode;
368                                 req->l_policy_data.l_flock.pid =
369                                         lock->l_policy_data.l_flock.pid;
370                                 req->l_policy_data.l_flock.start =
371                                         lock->l_policy_data.l_flock.start;
372                                 req->l_policy_data.l_flock.end =
373                                         lock->l_policy_data.l_flock.end;
374                                 *flags |= LDLM_FL_LOCK_CHANGED;
375                                 RETURN(LDLM_ITER_STOP);
376                         }
377
378                         /* add lock to blocking list before deadlock
379                          * check to prevent race */
380                         ldlm_flock_blocking_link(req, lock);
381
382                         if (ldlm_flock_deadlock(req, lock)) {
383                                 ldlm_flock_blocking_unlink(req);
384                                 ldlm_flock_destroy(req, mode, *flags);
385                                 *err = -EDEADLK;
386                                 RETURN(LDLM_ITER_STOP);
387                         }
388
389                         ldlm_resource_add_lock(res, &res->lr_waiting, req);
390                         *flags |= LDLM_FL_BLOCK_GRANTED;
391                         RETURN(LDLM_ITER_STOP);
392                 }
393                 if (reprocess_failed)
394                         RETURN(LDLM_ITER_CONTINUE);
395         }
396
397         if (*flags & LDLM_FL_TEST_LOCK) {
398                 ldlm_flock_destroy(req, mode, *flags);
399                 req->l_req_mode = LCK_NL;
400                 *flags |= LDLM_FL_LOCK_CHANGED;
401                 RETURN(LDLM_ITER_STOP);
402         }
403
404         /* In case we had slept on this lock request take it off of the
405          * deadlock detection hash list. */
406         ldlm_flock_blocking_unlink(req);
407
408         /* Scan the locks owned by this process that overlap this request.
409          * We may have to merge or split existing locks. */
410
411         if (!ownlocks)
412                 ownlocks = &res->lr_granted;
413
414         list_for_remaining_safe(ownlocks, tmp, &res->lr_granted) {
415                 lock = list_entry(ownlocks, struct ldlm_lock, l_res_link);
416
417                 if (!ldlm_same_flock_owner(lock, new))
418                         break;
419
420                 if (lock->l_granted_mode == mode) {
421                         /* If the modes are the same then we need to process
422                          * locks that overlap OR adjoin the new lock. The extra
423                          * logic condition is necessary to deal with arithmetic
424                          * overflow and underflow. */
425                         if ((new->l_policy_data.l_flock.start >
426                              (lock->l_policy_data.l_flock.end + 1))
427                             && (lock->l_policy_data.l_flock.end !=
428                                 OBD_OBJECT_EOF))
429                                 continue;
430
431                         if ((new->l_policy_data.l_flock.end <
432                              (lock->l_policy_data.l_flock.start - 1))
433                             && (lock->l_policy_data.l_flock.start != 0))
434                                 break;
435
436                         if (new->l_policy_data.l_flock.start <
437                             lock->l_policy_data.l_flock.start) {
438                                 lock->l_policy_data.l_flock.start =
439                                         new->l_policy_data.l_flock.start;
440                         } else {
441                                 new->l_policy_data.l_flock.start =
442                                         lock->l_policy_data.l_flock.start;
443                         }
444
445                         if (new->l_policy_data.l_flock.end >
446                             lock->l_policy_data.l_flock.end) {
447                                 lock->l_policy_data.l_flock.end =
448                                         new->l_policy_data.l_flock.end;
449                         } else {
450                                 new->l_policy_data.l_flock.end =
451                                         lock->l_policy_data.l_flock.end;
452                         }
453
454                         if (added) {
455                                 ldlm_flock_destroy(lock, mode, *flags);
456                         } else {
457                                 new = lock;
458                                 added = 1;
459                         }
460                         continue;
461                 }
462
463                 if (new->l_policy_data.l_flock.start >
464                     lock->l_policy_data.l_flock.end)
465                         continue;
466
467                 if (new->l_policy_data.l_flock.end <
468                     lock->l_policy_data.l_flock.start)
469                         break;
470
471                 ++overlaps;
472
473                 if (new->l_policy_data.l_flock.start <=
474                     lock->l_policy_data.l_flock.start) {
475                         if (new->l_policy_data.l_flock.end <
476                             lock->l_policy_data.l_flock.end) {
477                                 lock->l_policy_data.l_flock.start =
478                                         new->l_policy_data.l_flock.end + 1;
479                                 break;
480                         }
481                         ldlm_flock_destroy(lock, lock->l_req_mode, *flags);
482                         continue;
483                 }
484                 if (new->l_policy_data.l_flock.end >=
485                     lock->l_policy_data.l_flock.end) {
486                         lock->l_policy_data.l_flock.end =
487                                 new->l_policy_data.l_flock.start - 1;
488                         continue;
489                 }
490
491                 /* split the existing lock into two locks */
492
493                 /* if this is an F_UNLCK operation then we could avoid
494                  * allocating a new lock and use the req lock passed in
495                  * with the request but this would complicate the reply
496                  * processing since updates to req get reflected in the
497                  * reply. The client side replays the lock request so
498                  * it must see the original lock data in the reply. */
499
500                 /* XXX - if ldlm_lock_new() can sleep we should
501                  * release the lr_lock, allocate the new lock,
502                  * and restart processing this lock. */
503                 if (new2 == NULL) {
504                         unlock_res_and_lock(req);
505                         new2 = ldlm_lock_create(ns, &res->lr_name, LDLM_FLOCK,
506                                                 lock->l_granted_mode, &null_cbs,
507                                                 NULL, 0, LVB_T_NONE);
508                         lock_res_and_lock(req);
509                         if (IS_ERR(new2)) {
510                                 ldlm_flock_destroy(req, lock->l_granted_mode,
511                                                    *flags);
512                                 *err = PTR_ERR(new2);
513                                 RETURN(LDLM_ITER_STOP);
514                         }
515                         goto reprocess;
516                 }
517
518                 splitted = 1;
519
520                 new2->l_granted_mode = lock->l_granted_mode;
521                 new2->l_policy_data.l_flock.pid =
522                         new->l_policy_data.l_flock.pid;
523                 new2->l_policy_data.l_flock.owner =
524                         new->l_policy_data.l_flock.owner;
525                 new2->l_policy_data.l_flock.start =
526                         lock->l_policy_data.l_flock.start;
527                 new2->l_policy_data.l_flock.end =
528                         new->l_policy_data.l_flock.start - 1;
529                 lock->l_policy_data.l_flock.start =
530                         new->l_policy_data.l_flock.end + 1;
531                 new2->l_conn_export = lock->l_conn_export;
532                 if (lock->l_export != NULL) {
533                         new2->l_export = class_export_lock_get(lock->l_export, new2);
534                         if (new2->l_export->exp_lock_hash &&
535                             hlist_unhashed(&new2->l_exp_hash))
536                                 cfs_hash_add(new2->l_export->exp_lock_hash,
537                                              &new2->l_remote_handle,
538                                              &new2->l_exp_hash);
539                 }
540                 if (*flags == LDLM_FL_WAIT_NOREPROC)
541                         ldlm_lock_addref_internal_nolock(new2,
542                                                          lock->l_granted_mode);
543
544                 /* insert new2 at lock */
545                 ldlm_resource_add_lock(res, ownlocks, new2);
546                 LDLM_LOCK_RELEASE(new2);
547                 break;
548         }
549
550         /* if new2 is created but never used, destroy it*/
551         if (splitted == 0 && new2 != NULL)
552                 ldlm_lock_destroy_nolock(new2);
553
554         /* At this point we're granting the lock request. */
555         req->l_granted_mode = req->l_req_mode;
556
557         /* Add req to the granted queue before calling ldlm_reprocess_all(). */
558         if (!added) {
559                 list_del_init(&req->l_res_link);
560                 /* insert new lock before ownlocks in list. */
561                 ldlm_resource_add_lock(res, ownlocks, req);
562         }
563
564         if (*flags != LDLM_FL_WAIT_NOREPROC) {
565 #ifdef HAVE_SERVER_SUPPORT
566                 if (intention == LDLM_PROCESS_ENQUEUE) {
567                         /* If this is an unlock, reprocess the waitq and
568                          * send completions ASTs for locks that can now be
569                          * granted. The only problem with doing this
570                          * reprocessing here is that the completion ASTs for
571                          * newly granted locks will be sent before the unlock
572                          * completion is sent. It shouldn't be an issue. Also
573                          * note that ldlm_process_flock_lock() will recurse,
574                          * but only once because 'intention' won't be
575                          * LDLM_PROCESS_ENQUEUE from ldlm_reprocess_queue. */
576                         if ((mode == LCK_NL) && overlaps) {
577                                 struct list_head rpc_list;
578                                 int rc;
579
580                                 INIT_LIST_HEAD(&rpc_list);
581 restart:
582                                 ldlm_reprocess_queue(res, &res->lr_waiting,
583                                                      &rpc_list,
584                                                      LDLM_PROCESS_RESCAN);
585
586                                 unlock_res_and_lock(req);
587                                 rc = ldlm_run_ast_work(ns, &rpc_list,
588                                                        LDLM_WORK_CP_AST);
589                                 lock_res_and_lock(req);
590                                 if (rc == -ERESTART)
591                                         GOTO(restart, rc);
592                        }
593                 } else {
594                         LASSERT(req->l_completion_ast);
595                         ldlm_add_ast_work_item(req, NULL, grant_work);
596                 }
597 #else /* !HAVE_SERVER_SUPPORT */
598                 /* The only one possible case for client-side calls flock
599                  * policy function is ldlm_flock_completion_ast inside which
600                  * carries LDLM_FL_WAIT_NOREPROC flag. */
601                 CERROR("Illegal parameter for client-side-only module.\n");
602                 LBUG();
603 #endif /* HAVE_SERVER_SUPPORT */
604         }
605
606         /* In case we're reprocessing the requested lock we can't destroy
607          * it until after calling ldlm_add_ast_work_item() above so that laawi()
608          * can bump the reference count on \a req. Otherwise \a req
609          * could be freed before the completion AST can be sent.  */
610         if (added)
611                 ldlm_flock_destroy(req, mode, *flags);
612
613         ldlm_resource_dump(D_INFO, res);
614         RETURN(LDLM_ITER_CONTINUE);
615 }
616
617 struct ldlm_flock_wait_data {
618         struct ldlm_lock *fwd_lock;
619         int               fwd_generation;
620 };
621
622 static void
623 ldlm_flock_interrupted_wait(void *data)
624 {
625         struct ldlm_lock *lock;
626         ENTRY;
627
628         lock = ((struct ldlm_flock_wait_data *)data)->fwd_lock;
629
630         /* take lock off the deadlock detection hash list. */
631         lock_res_and_lock(lock);
632         ldlm_flock_blocking_unlink(lock);
633
634         /* client side - set flag to prevent lock from being put on LRU list */
635         ldlm_set_cbpending(lock);
636         unlock_res_and_lock(lock);
637
638         EXIT;
639 }
640
641 /**
642  * Flock completion callback function.
643  *
644  * \param lock [in,out]: A lock to be handled
645  * \param flags    [in]: flags
646  * \param *data    [in]: ldlm_work_cp_ast_lock() will use ldlm_cb_set_arg
647  *
648  * \retval 0    : success
649  * \retval <0   : failure
650  */
651 int
652 ldlm_flock_completion_ast(struct ldlm_lock *lock, __u64 flags, void *data)
653 {
654         struct file_lock *getlk = lock->l_ast_data;
655         struct obd_device *obd;
656         struct obd_import *imp = NULL;
657         struct ldlm_flock_wait_data fwd;
658         struct l_wait_info lwi;
659         enum ldlm_error err;
660         int rc = 0;
661         ENTRY;
662
663         OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_CP_CB_WAIT2, 4);
664         if (OBD_FAIL_PRECHECK(OBD_FAIL_LDLM_CP_CB_WAIT3)) {
665                 lock_res_and_lock(lock);
666                 lock->l_flags |= LDLM_FL_FAIL_LOC;
667                 unlock_res_and_lock(lock);
668                 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_CP_CB_WAIT3, 4);
669         }
670         CDEBUG(D_DLMTRACE, "flags: %#llx data: %p getlk: %p\n",
671                flags, data, getlk);
672
673         LASSERT(flags != LDLM_FL_WAIT_NOREPROC);
674
675         if (flags & LDLM_FL_FAILED)
676                 goto granted;
677
678         if (!(flags & LDLM_FL_BLOCKED_MASK)) {
679                 if (NULL == data)
680                         /* mds granted the lock in the reply */
681                         goto granted;
682                 /* CP AST RPC: lock get granted, wake it up */
683                 wake_up(&lock->l_waitq);
684                 RETURN(0);
685         }
686
687         LDLM_DEBUG(lock, "client-side enqueue returned a blocked lock, "
688                    "sleeping");
689         fwd.fwd_lock = lock;
690         obd = class_exp2obd(lock->l_conn_export);
691
692         /* if this is a local lock, there is no import */
693         if (NULL != obd)
694                 imp = obd->u.cli.cl_import;
695
696         if (NULL != imp) {
697                 spin_lock(&imp->imp_lock);
698                 fwd.fwd_generation = imp->imp_generation;
699                 spin_unlock(&imp->imp_lock);
700         }
701
702         lwi = LWI_TIMEOUT_INTR(0, NULL, ldlm_flock_interrupted_wait, &fwd);
703
704         /* Go to sleep until the lock is granted. */
705         rc = l_wait_event(lock->l_waitq, is_granted_or_cancelled(lock), &lwi);
706
707         if (rc) {
708                 LDLM_DEBUG(lock, "client-side enqueue waking up: failed (%d)",
709                            rc);
710                 RETURN(rc);
711         }
712
713 granted:
714         OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_CP_CB_WAIT, 10);
715
716         if (OBD_FAIL_PRECHECK(OBD_FAIL_LDLM_CP_CB_WAIT4)) {
717                 lock_res_and_lock(lock);
718                 /* DEADLOCK is always set with CBPENDING */
719                 lock->l_flags |= LDLM_FL_FLOCK_DEADLOCK | LDLM_FL_CBPENDING;
720                 unlock_res_and_lock(lock);
721                 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_CP_CB_WAIT4, 4);
722         }
723         if (OBD_FAIL_PRECHECK(OBD_FAIL_LDLM_CP_CB_WAIT5)) {
724                 lock_res_and_lock(lock);
725                 /* DEADLOCK is always set with CBPENDING */
726                 lock->l_flags |= LDLM_FL_FAIL_LOC |
727                                  LDLM_FL_FLOCK_DEADLOCK | LDLM_FL_CBPENDING;
728                 unlock_res_and_lock(lock);
729                 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_CP_CB_WAIT5, 4);
730         }
731
732         lock_res_and_lock(lock);
733
734
735         /* Protect against race where lock could have been just destroyed
736          * due to overlap in ldlm_process_flock_lock().
737          */
738         if (ldlm_is_destroyed(lock)) {
739                 unlock_res_and_lock(lock);
740                 LDLM_DEBUG(lock, "client-side enqueue waking up: destroyed");
741
742                 /* An error is still to be returned, to propagate it up to
743                  * ldlm_cli_enqueue_fini() caller. */
744                 RETURN(-EIO);
745         }
746
747         /* ldlm_lock_enqueue() has already placed lock on the granted list. */
748         ldlm_resource_unlink_lock(lock);
749
750         /* Import invalidation. We need to actually release the lock
751          * references being held, so that it can go away. No point in
752          * holding the lock even if app still believes it has it, since
753          * server already dropped it anyway. Only for granted locks too. */
754         /* Do the same for DEADLOCK'ed locks. */
755         if (ldlm_is_failed(lock) || ldlm_is_flock_deadlock(lock)) {
756                 int mode;
757
758                 if (flags & LDLM_FL_TEST_LOCK)
759                         LASSERT(ldlm_is_test_lock(lock));
760
761                 if (ldlm_is_test_lock(lock) || ldlm_is_flock_deadlock(lock))
762                         mode = getlk->fl_type;
763                 else
764                         mode = lock->l_granted_mode;
765
766                 if (ldlm_is_flock_deadlock(lock)) {
767                         LDLM_DEBUG(lock, "client-side enqueue deadlock "
768                                    "received");
769                         rc = -EDEADLK;
770                 }
771                 ldlm_flock_destroy(lock, mode, LDLM_FL_WAIT_NOREPROC);
772                 unlock_res_and_lock(lock);
773
774                 /* Need to wake up the waiter if we were evicted */
775                 wake_up(&lock->l_waitq);
776
777                 /* An error is still to be returned, to propagate it up to
778                  * ldlm_cli_enqueue_fini() caller. */
779                 RETURN(rc ? : -EIO);
780         }
781
782         LDLM_DEBUG(lock, "client-side enqueue granted");
783
784         if (flags & LDLM_FL_TEST_LOCK) {
785                 /*
786                  * fcntl(F_GETLK) request
787                  * The old mode was saved in getlk->fl_type so that if the mode
788                  * in the lock changes we can decref the appropriate refcount.
789                  */
790                 LASSERT(ldlm_is_test_lock(lock));
791                 ldlm_flock_destroy(lock, getlk->fl_type, LDLM_FL_WAIT_NOREPROC);
792                 switch (lock->l_granted_mode) {
793                 case LCK_PR:
794                         getlk->fl_type = F_RDLCK;
795                         break;
796                 case LCK_PW:
797                         getlk->fl_type = F_WRLCK;
798                         break;
799                 default:
800                         getlk->fl_type = F_UNLCK;
801                 }
802                 getlk->fl_pid = (pid_t)lock->l_policy_data.l_flock.pid;
803                 getlk->fl_start = (loff_t)lock->l_policy_data.l_flock.start;
804                 getlk->fl_end = (loff_t)lock->l_policy_data.l_flock.end;
805         } else {
806                 __u64 noreproc = LDLM_FL_WAIT_NOREPROC;
807
808                 /* We need to reprocess the lock to do merges or splits
809                  * with existing locks owned by this process. */
810                 ldlm_process_flock_lock(lock, &noreproc, 1, &err, NULL);
811         }
812         unlock_res_and_lock(lock);
813         RETURN(rc);
814 }
815 EXPORT_SYMBOL(ldlm_flock_completion_ast);
816
817 int ldlm_flock_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
818                             void *data, int flag)
819 {
820         ENTRY;
821
822         LASSERT(lock);
823         LASSERT(flag == LDLM_CB_CANCELING);
824
825         /* take lock off the deadlock detection hash list. */
826         lock_res_and_lock(lock);
827         ldlm_flock_blocking_unlink(lock);
828         unlock_res_and_lock(lock);
829         RETURN(0);
830 }
831
832 void ldlm_flock_policy_wire_to_local(const union ldlm_wire_policy_data *wpolicy,
833                                      union ldlm_policy_data *lpolicy)
834 {
835         lpolicy->l_flock.start = wpolicy->l_flock.lfw_start;
836         lpolicy->l_flock.end = wpolicy->l_flock.lfw_end;
837         lpolicy->l_flock.pid = wpolicy->l_flock.lfw_pid;
838         lpolicy->l_flock.owner = wpolicy->l_flock.lfw_owner;
839 }
840
841 void ldlm_flock_policy_local_to_wire(const union ldlm_policy_data *lpolicy,
842                                      union ldlm_wire_policy_data *wpolicy)
843 {
844         memset(wpolicy, 0, sizeof(*wpolicy));
845         wpolicy->l_flock.lfw_start = lpolicy->l_flock.start;
846         wpolicy->l_flock.lfw_end = lpolicy->l_flock.end;
847         wpolicy->l_flock.lfw_pid = lpolicy->l_flock.pid;
848         wpolicy->l_flock.lfw_owner = lpolicy->l_flock.owner;
849 }
850
851 /*
852  * Export handle<->flock hash operations.
853  */
854 static unsigned
855 ldlm_export_flock_hash(struct cfs_hash *hs, const void *key, unsigned mask)
856 {
857         return cfs_hash_u64_hash(*(__u64 *)key, mask);
858 }
859
860 static void *
861 ldlm_export_flock_key(struct hlist_node *hnode)
862 {
863         struct ldlm_lock *lock;
864
865         lock = hlist_entry(hnode, struct ldlm_lock, l_exp_flock_hash);
866         return &lock->l_policy_data.l_flock.owner;
867 }
868
869 static int
870 ldlm_export_flock_keycmp(const void *key, struct hlist_node *hnode)
871 {
872         return !memcmp(ldlm_export_flock_key(hnode), key, sizeof(__u64));
873 }
874
875 static void *
876 ldlm_export_flock_object(struct hlist_node *hnode)
877 {
878         return hlist_entry(hnode, struct ldlm_lock, l_exp_flock_hash);
879 }
880
881 static void
882 ldlm_export_flock_get(struct cfs_hash *hs, struct hlist_node *hnode)
883 {
884         struct ldlm_lock *lock;
885         struct ldlm_flock *flock;
886
887         lock = hlist_entry(hnode, struct ldlm_lock, l_exp_flock_hash);
888         LDLM_LOCK_GET(lock);
889
890         flock = &lock->l_policy_data.l_flock;
891         LASSERT(flock->blocking_export != NULL);
892         class_export_get(flock->blocking_export);
893         atomic_inc(&flock->blocking_refs);
894 }
895
896 static void
897 ldlm_export_flock_put(struct cfs_hash *hs, struct hlist_node *hnode)
898 {
899         struct ldlm_lock *lock;
900         struct ldlm_flock *flock;
901
902         lock = hlist_entry(hnode, struct ldlm_lock, l_exp_flock_hash);
903
904         flock = &lock->l_policy_data.l_flock;
905         LASSERT(flock->blocking_export != NULL);
906         class_export_put(flock->blocking_export);
907         if (atomic_dec_and_test(&flock->blocking_refs)) {
908                 flock->blocking_owner = 0;
909                 flock->blocking_export = NULL;
910         }
911         LDLM_LOCK_RELEASE(lock);
912 }
913
914 static struct cfs_hash_ops ldlm_export_flock_ops = {
915         .hs_hash        = ldlm_export_flock_hash,
916         .hs_key         = ldlm_export_flock_key,
917         .hs_keycmp      = ldlm_export_flock_keycmp,
918         .hs_object      = ldlm_export_flock_object,
919         .hs_get         = ldlm_export_flock_get,
920         .hs_put         = ldlm_export_flock_put,
921         .hs_put_locked  = ldlm_export_flock_put,
922 };
923
924 int ldlm_init_flock_export(struct obd_export *exp)
925 {
926         if( strcmp(exp->exp_obd->obd_type->typ_name, LUSTRE_MDT_NAME) != 0)
927                 RETURN(0);
928
929         exp->exp_flock_hash =
930                 cfs_hash_create(obd_uuid2str(&exp->exp_client_uuid),
931                                 HASH_EXP_LOCK_CUR_BITS,
932                                 HASH_EXP_LOCK_MAX_BITS,
933                                 HASH_EXP_LOCK_BKT_BITS, 0,
934                                 CFS_HASH_MIN_THETA, CFS_HASH_MAX_THETA,
935                                 &ldlm_export_flock_ops,
936                                 CFS_HASH_DEFAULT | CFS_HASH_NBLK_CHANGE);
937         if (!exp->exp_flock_hash)
938                 RETURN(-ENOMEM);
939
940         RETURN(0);
941 }
942
943 void ldlm_destroy_flock_export(struct obd_export *exp)
944 {
945         ENTRY;
946         if (exp->exp_flock_hash) {
947                 cfs_hash_putref(exp->exp_flock_hash);
948                 exp->exp_flock_hash = NULL;
949         }
950         EXIT;
951 }