Whamcloud - gitweb
LU-11158 mdt: grow lvb buffer to hold layout
[fs/lustre-release.git] / lustre / ldlm / ldlm_lockd.c
1 /*
2  * GPL HEADER START
3  *
4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
5  *
6  * This program is free software; you can redistribute it and/or modify
7  * it under the terms of the GNU General Public License version 2 only,
8  * as published by the Free Software Foundation.
9  *
10  * This program is distributed in the hope that it will be useful, but
11  * WITHOUT ANY WARRANTY; without even the implied warranty of
12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
13  * General Public License version 2 for more details (a copy is included
14  * in the LICENSE file that accompanied this code).
15  *
16  * You should have received a copy of the GNU General Public License
17  * version 2 along with this program; If not, see
18  * http://www.gnu.org/licenses/gpl-2.0.html
19  *
20  * GPL HEADER END
21  */
22 /*
23  * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
24  * Use is subject to license terms.
25  *
26  * Copyright (c) 2010, 2017, Intel Corporation.
27  */
28 /*
29  * This file is part of Lustre, http://www.lustre.org/
30  * Lustre is a trademark of Sun Microsystems, Inc.
31  *
32  * lustre/ldlm/ldlm_lockd.c
33  *
34  * Author: Peter Braam <braam@clusterfs.com>
35  * Author: Phil Schwan <phil@clusterfs.com>
36  */
37
38 #define DEBUG_SUBSYSTEM S_LDLM
39
40 #include <linux/kthread.h>
41 #include <linux/list.h>
42 #include <libcfs/libcfs.h>
43 #include <lustre_errno.h>
44 #include <lustre_dlm.h>
45 #include <obd_class.h>
46 #include "ldlm_internal.h"
47
48 static int ldlm_num_threads;
49 module_param(ldlm_num_threads, int, 0444);
50 MODULE_PARM_DESC(ldlm_num_threads, "number of DLM service threads to start");
51
52 static unsigned int ldlm_cpu_bind = 1;
53 module_param(ldlm_cpu_bind, uint, 0444);
54 MODULE_PARM_DESC(ldlm_cpu_bind,
55                  "bind DLM service threads to particular CPU partitions");
56
57 static char *ldlm_cpts;
58 module_param(ldlm_cpts, charp, 0444);
59 MODULE_PARM_DESC(ldlm_cpts, "CPU partitions ldlm threads should run on");
60
61 static DEFINE_MUTEX(ldlm_ref_mutex);
62 static int ldlm_refcount;
63
64 struct kobject *ldlm_kobj;
65 struct kset *ldlm_ns_kset;
66 struct kset *ldlm_svc_kset;
67
68 /* LDLM state */
69
70 static struct ldlm_state *ldlm_state;
71
72 /* timeout for initial callback (AST) reply (bz10399)
73  * Due to having to send a 32 bit time value over the
74  * wire return it as time_t instead of time64_t
75  */
76 static inline time_t ldlm_get_rq_timeout(void)
77 {
78         /* Non-AT value */
79         time_t timeout = min(ldlm_timeout, obd_timeout / 3);
80
81         return timeout < 1 ? 1 : timeout;
82 }
83
84 struct ldlm_bl_pool {
85         spinlock_t              blp_lock;
86
87         /*
88          * blp_prio_list is used for callbacks that should be handled
89          * as a priority. It is used for LDLM_FL_DISCARD_DATA requests.
90          * see bug 13843
91          */
92         struct list_head              blp_prio_list;
93
94         /*
95          * blp_list is used for all other callbacks which are likely
96          * to take longer to process.
97          */
98         struct list_head              blp_list;
99
100         wait_queue_head_t       blp_waitq;
101         struct completion       blp_comp;
102         atomic_t            blp_num_threads;
103         atomic_t            blp_busy_threads;
104         int                     blp_min_threads;
105         int                     blp_max_threads;
106 };
107
108 struct ldlm_bl_work_item {
109         struct list_head        blwi_entry;
110         struct ldlm_namespace   *blwi_ns;
111         struct ldlm_lock_desc   blwi_ld;
112         struct ldlm_lock        *blwi_lock;
113         struct list_head        blwi_head;
114         int                     blwi_count;
115         struct completion       blwi_comp;
116         enum ldlm_cancel_flags  blwi_flags;
117         int                     blwi_mem_pressure;
118 };
119
120 #ifdef HAVE_SERVER_SUPPORT
121
122 /**
123  * Protects both waiting_locks_list and expired_lock_thread.
124  */
125 static DEFINE_SPINLOCK(waiting_locks_spinlock); /* BH lock (timer) */
126
127 /**
128  * List for contended locks.
129  *
130  * As soon as a lock is contended, it gets placed on this list and
131  * expected time to get a response is filled in the lock. A special
132  * thread walks the list looking for locks that should be released and
133  * schedules client evictions for those that have not been released in
134  * time.
135  *
136  * All access to it should be under waiting_locks_spinlock.
137  */
138 static LIST_HEAD(waiting_locks_list);
139 static void waiting_locks_callback(TIMER_DATA_TYPE unused);
140 static CFS_DEFINE_TIMER(waiting_locks_timer, waiting_locks_callback, 0, 0);
141
142 enum elt_state {
143         ELT_STOPPED,
144         ELT_READY,
145         ELT_TERMINATE,
146 };
147
148 static DECLARE_WAIT_QUEUE_HEAD(expired_lock_wait_queue);
149 static enum elt_state expired_lock_thread_state = ELT_STOPPED;
150 static int expired_lock_dump;
151 static LIST_HEAD(expired_lock_list);
152
153 static int ldlm_lock_busy(struct ldlm_lock *lock);
154 static int ldlm_add_waiting_lock(struct ldlm_lock *lock, time64_t timeout);
155 static int __ldlm_add_waiting_lock(struct ldlm_lock *lock, time64_t seconds);
156
157 static inline int have_expired_locks(void)
158 {
159         int need_to_run;
160
161         ENTRY;
162         spin_lock_bh(&waiting_locks_spinlock);
163         need_to_run = !list_empty(&expired_lock_list);
164         spin_unlock_bh(&waiting_locks_spinlock);
165
166         RETURN(need_to_run);
167 }
168
169 /**
170  * Check expired lock list for expired locks and time them out.
171  */
172 static int expired_lock_main(void *arg)
173 {
174         struct list_head *expired = &expired_lock_list;
175         struct l_wait_info lwi = { 0 };
176         int do_dump;
177
178         ENTRY;
179
180         expired_lock_thread_state = ELT_READY;
181         wake_up(&expired_lock_wait_queue);
182
183         while (1) {
184                 l_wait_event(expired_lock_wait_queue,
185                              have_expired_locks() ||
186                              expired_lock_thread_state == ELT_TERMINATE,
187                              &lwi);
188
189                 spin_lock_bh(&waiting_locks_spinlock);
190                 if (expired_lock_dump) {
191                         spin_unlock_bh(&waiting_locks_spinlock);
192
193                         /* from waiting_locks_callback, but not in timer */
194                         libcfs_debug_dumplog();
195
196                         spin_lock_bh(&waiting_locks_spinlock);
197                         expired_lock_dump = 0;
198                 }
199
200                 do_dump = 0;
201
202                 while (!list_empty(expired)) {
203                         struct obd_export *export;
204                         struct ldlm_lock *lock;
205
206                         lock = list_entry(expired->next, struct ldlm_lock,
207                                           l_pending_chain);
208                         if ((void *)lock < LP_POISON + PAGE_SIZE &&
209                             (void *)lock >= LP_POISON) {
210                                 spin_unlock_bh(&waiting_locks_spinlock);
211                                 CERROR("free lock on elt list %p\n", lock);
212                                 LBUG();
213                         }
214                         list_del_init(&lock->l_pending_chain);
215                         if ((void *)lock->l_export <
216                              LP_POISON + PAGE_SIZE &&
217                             (void *)lock->l_export >= LP_POISON) {
218                                 CERROR("lock with free export on elt list %p\n",
219                                        lock->l_export);
220                                 lock->l_export = NULL;
221                                 LDLM_ERROR(lock, "free export");
222                                 /* release extra ref grabbed by
223                                  * ldlm_add_waiting_lock() or
224                                  * ldlm_failed_ast() */
225                                 LDLM_LOCK_RELEASE(lock);
226                                 continue;
227                         }
228
229                         if (ldlm_is_destroyed(lock)) {
230                                 /* release the lock refcount where
231                                  * waiting_locks_callback() founds */
232                                 LDLM_LOCK_RELEASE(lock);
233                                 continue;
234                         }
235                         export = class_export_lock_get(lock->l_export, lock);
236                         spin_unlock_bh(&waiting_locks_spinlock);
237
238                         /* Check if we need to prolong timeout */
239                         if (!OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_TIMEOUT) &&
240                             lock->l_callback_timeout != 0 && /* not AST error */
241                             ldlm_lock_busy(lock)) {
242                                 LDLM_DEBUG(lock, "prolong the busy lock");
243                                 lock_res_and_lock(lock);
244                                 ldlm_add_waiting_lock(lock,
245                                                 ldlm_bl_timeout(lock) >> 1);
246                                 unlock_res_and_lock(lock);
247                         } else {
248                                 spin_lock_bh(&export->exp_bl_list_lock);
249                                 list_del_init(&lock->l_exp_list);
250                                 spin_unlock_bh(&export->exp_bl_list_lock);
251
252                                 LDLM_ERROR(lock,
253                                            "lock callback timer expired after %llds: evicting client at %s ",
254                                            ktime_get_real_seconds() -
255                                            lock->l_blast_sent,
256                                            obd_export_nid2str(export));
257                                 ldlm_lock_to_ns(lock)->ns_timeouts++;
258                                 do_dump++;
259                                 class_fail_export(export);
260                         }
261                         class_export_lock_put(export, lock);
262                         /* release extra ref grabbed by ldlm_add_waiting_lock()
263                          * or ldlm_failed_ast() */
264                         LDLM_LOCK_RELEASE(lock);
265
266                         spin_lock_bh(&waiting_locks_spinlock);
267                 }
268                 spin_unlock_bh(&waiting_locks_spinlock);
269
270                 if (do_dump && obd_dump_on_eviction) {
271                         CERROR("dump the log upon eviction\n");
272                         libcfs_debug_dumplog();
273                 }
274
275                 if (expired_lock_thread_state == ELT_TERMINATE)
276                         break;
277         }
278
279         expired_lock_thread_state = ELT_STOPPED;
280         wake_up(&expired_lock_wait_queue);
281         RETURN(0);
282 }
283
284 /**
285  * Check if there is a request in the export request list
286  * which prevents the lock canceling.
287  */
288 static int ldlm_lock_busy(struct ldlm_lock *lock)
289 {
290         struct ptlrpc_request *req;
291         int match = 0;
292         ENTRY;
293
294         if (lock->l_export == NULL)
295                 return 0;
296
297         spin_lock(&lock->l_export->exp_rpc_lock);
298         list_for_each_entry(req, &lock->l_export->exp_hp_rpcs,
299                                 rq_exp_list) {
300                 if (req->rq_ops->hpreq_lock_match) {
301                         match = req->rq_ops->hpreq_lock_match(req, lock);
302                         if (match)
303                                 break;
304                 }
305         }
306         spin_unlock(&lock->l_export->exp_rpc_lock);
307         RETURN(match);
308 }
309
310 /* This is called from within a timer interrupt and cannot schedule */
311 static void waiting_locks_callback(TIMER_DATA_TYPE unused)
312 {
313         struct ldlm_lock        *lock;
314         int                     need_dump = 0;
315
316         spin_lock_bh(&waiting_locks_spinlock);
317         while (!list_empty(&waiting_locks_list)) {
318                 lock = list_entry(waiting_locks_list.next, struct ldlm_lock,
319                                   l_pending_chain);
320                 if (lock->l_callback_timeout > ktime_get_seconds() ||
321                     lock->l_req_mode == LCK_GROUP)
322                         break;
323
324                 /* no needs to take an extra ref on the lock since it was in
325                  * the waiting_locks_list and ldlm_add_waiting_lock()
326                  * already grabbed a ref */
327                 list_del(&lock->l_pending_chain);
328                 list_add(&lock->l_pending_chain, &expired_lock_list);
329                 need_dump = 1;
330         }
331
332         if (!list_empty(&expired_lock_list)) {
333                 if (obd_dump_on_timeout && need_dump)
334                         expired_lock_dump = __LINE__;
335
336                 wake_up(&expired_lock_wait_queue);
337         }
338
339         /*
340          * Make sure the timer will fire again if we have any locks
341          * left.
342          */
343         if (!list_empty(&waiting_locks_list)) {
344                 unsigned long timeout_jiffies;
345
346                 lock = list_entry(waiting_locks_list.next, struct ldlm_lock,
347                                   l_pending_chain);
348                 timeout_jiffies = cfs_time_seconds(lock->l_callback_timeout);
349                 mod_timer(&waiting_locks_timer, timeout_jiffies);
350         }
351         spin_unlock_bh(&waiting_locks_spinlock);
352 }
353
354 /**
355  * Add lock to the list of contended locks.
356  *
357  * Indicate that we're waiting for a client to call us back cancelling a given
358  * lock.  We add it to the pending-callback chain, and schedule the lock-timeout
359  * timer to fire appropriately.  (We round up to the next second, to avoid
360  * floods of timer firings during periods of high lock contention and traffic).
361  * As done by ldlm_add_waiting_lock(), the caller must grab a lock reference
362  * if it has been added to the waiting list (1 is returned).
363  *
364  * Called with the namespace lock held.
365  */
366 static int __ldlm_add_waiting_lock(struct ldlm_lock *lock, time64_t seconds)
367 {
368         unsigned long timeout_jiffies;
369         time64_t timeout;
370
371         if (!list_empty(&lock->l_pending_chain))
372                 return 0;
373
374         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_NOTIMEOUT) ||
375             OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_TIMEOUT))
376                 seconds = 1;
377
378         timeout = ktime_get_seconds() + seconds;
379         if (likely(timeout > lock->l_callback_timeout))
380                 lock->l_callback_timeout = timeout;
381
382         timeout_jiffies = cfs_time_seconds(lock->l_callback_timeout);
383
384         if (time_before(timeout_jiffies, waiting_locks_timer.expires) ||
385             !timer_pending(&waiting_locks_timer))
386                 mod_timer(&waiting_locks_timer, timeout_jiffies);
387
388         /* if the new lock has a shorter timeout than something earlier on
389          * the list, we'll wait the longer amount of time; no big deal.
390          */
391         /* FIFO */
392         list_add_tail(&lock->l_pending_chain, &waiting_locks_list);
393         return 1;
394 }
395
396 static void ldlm_add_blocked_lock(struct ldlm_lock *lock)
397 {
398         spin_lock_bh(&lock->l_export->exp_bl_list_lock);
399         if (list_empty(&lock->l_exp_list)) {
400                 if (lock->l_granted_mode != lock->l_req_mode)
401                         list_add_tail(&lock->l_exp_list,
402                                       &lock->l_export->exp_bl_list);
403                 else
404                         list_add(&lock->l_exp_list,
405                                  &lock->l_export->exp_bl_list);
406         }
407         spin_unlock_bh(&lock->l_export->exp_bl_list_lock);
408
409         /* A blocked lock is added. Adjust the position in
410          * the stale list if the export is in the list.
411          * If export is stale and not in the list - it is being
412          * processed and will be placed on the right position
413          * on obd_stale_export_put(). */
414         if (!list_empty(&lock->l_export->exp_stale_list))
415                 obd_stale_export_adjust(lock->l_export);
416 }
417
418 static int ldlm_add_waiting_lock(struct ldlm_lock *lock, time64_t timeout)
419 {
420         int ret;
421
422         /* NB: must be called with hold of lock_res_and_lock() */
423         LASSERT(ldlm_is_res_locked(lock));
424         LASSERT(!ldlm_is_cancel_on_block(lock));
425
426         /* Do not put cross-MDT lock in the waiting list, since we
427          * will not evict it due to timeout for now */
428         if (lock->l_export != NULL &&
429             (exp_connect_flags(lock->l_export) & OBD_CONNECT_MDS_MDS))
430                 return 0;
431
432         spin_lock_bh(&waiting_locks_spinlock);
433         if (ldlm_is_cancel(lock)) {
434                 spin_unlock_bh(&waiting_locks_spinlock);
435                 return 0;
436         }
437
438         if (ldlm_is_destroyed(lock)) {
439                 static time64_t next;
440
441                 spin_unlock_bh(&waiting_locks_spinlock);
442                 LDLM_ERROR(lock, "not waiting on destroyed lock (bug 5653)");
443                 if (ktime_get_seconds() > next) {
444                         next = ktime_get_seconds() + 14400;
445                         libcfs_debug_dumpstack(NULL);
446                 }
447                 return 0;
448         }
449
450         ldlm_set_waited(lock);
451         lock->l_blast_sent = ktime_get_real_seconds();
452         ret = __ldlm_add_waiting_lock(lock, timeout);
453         if (ret) {
454                 /* grab ref on the lock if it has been added to the
455                  * waiting list */
456                 LDLM_LOCK_GET(lock);
457         }
458         spin_unlock_bh(&waiting_locks_spinlock);
459
460         if (ret)
461                 ldlm_add_blocked_lock(lock);
462
463         LDLM_DEBUG(lock, "%sadding to wait list(timeout: %lld, AT: %s)",
464                    ret == 0 ? "not re-" : "", timeout,
465                    AT_OFF ? "off" : "on");
466         return ret;
467 }
468
469 /**
470  * Remove a lock from the pending list, likely because it had its cancellation
471  * callback arrive without incident.  This adjusts the lock-timeout timer if
472  * needed.  Returns 0 if the lock wasn't pending after all, 1 if it was.
473  * As done by ldlm_del_waiting_lock(), the caller must release the lock
474  * reference when the lock is removed from any list (1 is returned).
475  *
476  * Called with namespace lock held.
477  */
478 static int __ldlm_del_waiting_lock(struct ldlm_lock *lock)
479 {
480         struct list_head *list_next;
481
482         if (list_empty(&lock->l_pending_chain))
483                 return 0;
484
485         list_next = lock->l_pending_chain.next;
486         if (lock->l_pending_chain.prev == &waiting_locks_list) {
487                 /* Removing the head of the list, adjust timer. */
488                 if (list_next == &waiting_locks_list) {
489                         /* No more, just cancel. */
490                         del_timer(&waiting_locks_timer);
491                 } else {
492                         struct ldlm_lock *next;
493
494                         next = list_entry(list_next, struct ldlm_lock,
495                                           l_pending_chain);
496                         mod_timer(&waiting_locks_timer,
497                                   cfs_time_seconds(next->l_callback_timeout));
498                 }
499         }
500         list_del_init(&lock->l_pending_chain);
501
502         return 1;
503 }
504
505 int ldlm_del_waiting_lock(struct ldlm_lock *lock)
506 {
507         int ret;
508
509         if (lock->l_export == NULL) {
510                 /* We don't have a "waiting locks list" on clients. */
511                 CDEBUG(D_DLMTRACE, "Client lock %p : no-op\n", lock);
512                 return 0;
513         }
514
515         spin_lock_bh(&waiting_locks_spinlock);
516         ret = __ldlm_del_waiting_lock(lock);
517         ldlm_clear_waited(lock);
518         spin_unlock_bh(&waiting_locks_spinlock);
519
520         /* remove the lock out of export blocking list */
521         spin_lock_bh(&lock->l_export->exp_bl_list_lock);
522         list_del_init(&lock->l_exp_list);
523         spin_unlock_bh(&lock->l_export->exp_bl_list_lock);
524
525         if (ret) {
526                 /* release lock ref if it has indeed been removed
527                  * from a list */
528                 LDLM_LOCK_RELEASE(lock);
529         }
530
531         LDLM_DEBUG(lock, "%s", ret == 0 ? "wasn't waiting" : "removed");
532         return ret;
533 }
534
535 /**
536  * Prolong the contended lock waiting time.
537  *
538  * Called with namespace lock held.
539  */
540 int ldlm_refresh_waiting_lock(struct ldlm_lock *lock, time64_t timeout)
541 {
542         if (lock->l_export == NULL) {
543                 /* We don't have a "waiting locks list" on clients. */
544                 LDLM_DEBUG(lock, "client lock: no-op");
545                 return 0;
546         }
547
548         if (exp_connect_flags(lock->l_export) & OBD_CONNECT_MDS_MDS) {
549                 /* We don't have a "waiting locks list" on OSP. */
550                 LDLM_DEBUG(lock, "MDS-MDS lock: no-op");
551                 return 0;
552         }
553
554         spin_lock_bh(&waiting_locks_spinlock);
555
556         if (list_empty(&lock->l_pending_chain)) {
557                 spin_unlock_bh(&waiting_locks_spinlock);
558                 LDLM_DEBUG(lock, "wasn't waiting");
559                 return 0;
560         }
561
562         /* we remove/add the lock to the waiting list, so no needs to
563          * release/take a lock reference */
564         __ldlm_del_waiting_lock(lock);
565         __ldlm_add_waiting_lock(lock, timeout);
566         spin_unlock_bh(&waiting_locks_spinlock);
567
568         LDLM_DEBUG(lock, "refreshed");
569         return 1;
570 }
571 EXPORT_SYMBOL(ldlm_refresh_waiting_lock);
572
573 #else /* HAVE_SERVER_SUPPORT */
574
575 int ldlm_del_waiting_lock(struct ldlm_lock *lock)
576 {
577         RETURN(0);
578 }
579
580 int ldlm_refresh_waiting_lock(struct ldlm_lock *lock, time64_t timeout)
581 {
582         RETURN(0);
583 }
584
585 #endif /* !HAVE_SERVER_SUPPORT */
586
587 #ifdef HAVE_SERVER_SUPPORT
588
589 /**
590  * Calculate the per-export Blocking timeout (covering BL AST, data flush,
591  * lock cancel, and their replies). Used for lock callback timeout and AST
592  * re-send period.
593  *
594  * \param[in] lock        lock which is getting the blocking callback
595  *
596  * \retval            timeout in seconds to wait for the client reply
597  */
598 time64_t ldlm_bl_timeout(struct ldlm_lock *lock)
599 {
600         time64_t timeout;
601
602         if (AT_OFF)
603                 return obd_timeout / 2;
604
605         /* Since these are non-updating timeouts, we should be conservative.
606          * Take more than usually, 150%
607          * It would be nice to have some kind of "early reply" mechanism for
608          * lock callbacks too... */
609         timeout = at_get(&lock->l_export->exp_bl_lock_at);
610         return max(timeout + (timeout >> 1), (time64_t)ldlm_enqueue_min);
611 }
612 EXPORT_SYMBOL(ldlm_bl_timeout);
613
614 /**
615  * Perform lock cleanup if AST sending failed.
616  */
617 static void ldlm_failed_ast(struct ldlm_lock *lock, int rc,
618                             const char *ast_type)
619 {
620         LCONSOLE_ERROR_MSG(0x138, "%s: A client on nid %s was evicted due "
621                            "to a lock %s callback time out: rc %d\n",
622                            lock->l_export->exp_obd->obd_name,
623                            obd_export_nid2str(lock->l_export), ast_type, rc);
624
625         if (obd_dump_on_timeout)
626                 libcfs_debug_dumplog();
627         spin_lock_bh(&waiting_locks_spinlock);
628         if (__ldlm_del_waiting_lock(lock) == 0)
629                 /* the lock was not in any list, grab an extra ref before adding
630                  * the lock to the expired list */
631                 LDLM_LOCK_GET(lock);
632         lock->l_callback_timeout = 0; /* differentiate it from expired locks */
633         list_add(&lock->l_pending_chain, &expired_lock_list);
634         wake_up(&expired_lock_wait_queue);
635         spin_unlock_bh(&waiting_locks_spinlock);
636 }
637
638 /**
639  * Perform lock cleanup if AST reply came with error.
640  */
641 static int ldlm_handle_ast_error(const struct lu_env *env,
642                                  struct ldlm_lock *lock,
643                                  struct ptlrpc_request *req, int rc,
644                                  const char *ast_type)
645 {
646         struct lnet_process_id peer = req->rq_import->imp_connection->c_peer;
647
648         if (!req->rq_replied || (rc && rc != -EINVAL)) {
649                 if (ldlm_is_cancel(lock)) {
650                         LDLM_DEBUG(lock,
651                                    "%s AST (req@%p x%llu) timeout from nid %s, but cancel was received (AST reply lost?)",
652                                    ast_type, req, req->rq_xid,
653                                    libcfs_nid2str(peer.nid));
654                         ldlm_lock_cancel(lock);
655                         rc = -ERESTART;
656                 } else if (rc == -ENODEV || rc == -ESHUTDOWN ||
657                            (rc == -EIO &&
658                             req->rq_import->imp_state == LUSTRE_IMP_CLOSED)) {
659                         /* Upon umount process the AST fails because cannot be
660                          * sent. This shouldn't lead to the client eviction.
661                          * -ENODEV error is returned by ptl_send_rpc() for
662                          *  new request in such import.
663                          * -SHUTDOWN is returned by ptlrpc_import_delay_req()
664                          *  if imp_invalid is set or obd_no_recov.
665                          * Meanwhile there is also check for LUSTRE_IMP_CLOSED
666                          * in ptlrpc_import_delay_req() as well with -EIO code.
667                          * In all such cases errors are ignored.
668                          */
669                         LDLM_DEBUG(lock, "%s AST can't be sent due to a server"
670                                          " %s failure or umount process: rc = %d\n",
671                                          ast_type,
672                                          req->rq_import->imp_obd->obd_name, rc);
673                 } else {
674                         LDLM_ERROR(lock,
675                                    "client (nid %s) %s %s AST (req@%p x%llu status %d rc %d), evict it",
676                                    libcfs_nid2str(peer.nid),
677                                    req->rq_replied ? "returned error from" :
678                                    "failed to reply to",
679                                    ast_type, req, req->rq_xid,
680                                    (req->rq_repmsg != NULL) ?
681                                    lustre_msg_get_status(req->rq_repmsg) : 0,
682                                    rc);
683                         ldlm_failed_ast(lock, rc, ast_type);
684                 }
685                 return rc;
686         }
687
688         if (rc == -EINVAL) {
689                 struct ldlm_resource *res = lock->l_resource;
690
691                 LDLM_DEBUG(lock,
692                            "client (nid %s) returned %d from %s AST (req@%p x%llu) - normal race",
693                            libcfs_nid2str(peer.nid),
694                            req->rq_repmsg ?
695                            lustre_msg_get_status(req->rq_repmsg) : -1,
696                            ast_type, req, req->rq_xid);
697                 if (res) {
698                         /* update lvbo to return proper attributes.
699                          * see bug 23174 */
700                         ldlm_resource_getref(res);
701                         ldlm_lvbo_update(env, res, lock, NULL, 1);
702                         ldlm_resource_putref(res);
703                 }
704                 ldlm_lock_cancel(lock);
705                 rc = -ERESTART;
706         }
707
708         return rc;
709 }
710
711 static int ldlm_cb_interpret(const struct lu_env *env,
712                              struct ptlrpc_request *req, void *data, int rc)
713 {
714         struct ldlm_cb_async_args *ca   = data;
715         struct ldlm_lock          *lock = ca->ca_lock;
716         struct ldlm_cb_set_arg    *arg  = ca->ca_set_arg;
717         ENTRY;
718
719         LASSERT(lock != NULL);
720
721         switch (arg->type) {
722         case LDLM_GL_CALLBACK:
723                 /* Update the LVB from disk if the AST failed
724                  * (this is a legal race)
725                  *
726                  * - Glimpse callback of local lock just returns
727                  *   -ELDLM_NO_LOCK_DATA.
728                  * - Glimpse callback of remote lock might return
729                  *   -ELDLM_NO_LOCK_DATA when inode is cleared. LU-274
730                  */
731                 if (unlikely(arg->gl_interpret_reply)) {
732                         rc = arg->gl_interpret_reply(env, req, data, rc);
733                 } else if (rc == -ELDLM_NO_LOCK_DATA) {
734                         LDLM_DEBUG(lock, "lost race - client has a lock but no "
735                                    "inode");
736                         ldlm_lvbo_update(env, lock->l_resource, lock, NULL, 1);
737                 } else if (rc != 0) {
738                         rc = ldlm_handle_ast_error(env, lock, req,
739                                                    rc, "glimpse");
740                 } else {
741                         rc = ldlm_lvbo_update(env, lock->l_resource,
742                                               lock, req, 1);
743                 }
744                 break;
745         case LDLM_BL_CALLBACK:
746                 if (rc != 0)
747                         rc = ldlm_handle_ast_error(env, lock, req,
748                                                    rc, "blocking");
749                 break;
750         case LDLM_CP_CALLBACK:
751                 if (rc != 0)
752                         rc = ldlm_handle_ast_error(env, lock, req,
753                                                    rc, "completion");
754                 break;
755         default:
756                 LDLM_ERROR(lock, "invalid opcode for lock callback %d",
757                            arg->type);
758                 LBUG();
759         }
760
761         /* release extra reference taken in ldlm_ast_fini() */
762         LDLM_LOCK_RELEASE(lock);
763
764         if (rc == -ERESTART)
765                 atomic_inc(&arg->restart);
766
767         RETURN(0);
768 }
769
770 static void ldlm_update_resend(struct ptlrpc_request *req, void *data)
771 {
772         struct ldlm_cb_async_args *ca = data;
773         struct ldlm_lock *lock = ca->ca_lock;
774
775         ldlm_refresh_waiting_lock(lock, ldlm_bl_timeout(lock));
776 }
777
778 static inline int ldlm_ast_fini(struct ptlrpc_request *req,
779                                 struct ldlm_cb_set_arg *arg,
780                                 struct ldlm_lock *lock,
781                                 int instant_cancel)
782 {
783         int rc = 0;
784         ENTRY;
785
786         if (unlikely(instant_cancel)) {
787                 rc = ptl_send_rpc(req, 1);
788                 ptlrpc_req_finished(req);
789                 if (rc == 0)
790                         atomic_inc(&arg->restart);
791         } else {
792                 LDLM_LOCK_GET(lock);
793                 ptlrpc_set_add_req(arg->set, req);
794         }
795
796         RETURN(rc);
797 }
798
799 /**
800  * Check if there are requests in the export request list which prevent
801  * the lock canceling and make these requests high priority ones.
802  */
803 static void ldlm_lock_reorder_req(struct ldlm_lock *lock)
804 {
805         struct ptlrpc_request *req;
806         ENTRY;
807
808         if (lock->l_export == NULL) {
809                 LDLM_DEBUG(lock, "client lock: no-op");
810                 RETURN_EXIT;
811         }
812
813         spin_lock(&lock->l_export->exp_rpc_lock);
814         list_for_each_entry(req, &lock->l_export->exp_hp_rpcs,
815                             rq_exp_list) {
816                 /* Do not process requests that were not yet added to there
817                  * incoming queue or were already removed from there for
818                  * processing. We evaluate ptlrpc_nrs_req_can_move() without
819                  * holding svcpt->scp_req_lock, and then redo the check with
820                  * the lock held once we need to obtain a reliable result.
821                  */
822                 if (ptlrpc_nrs_req_can_move(req) &&
823                     req->rq_ops->hpreq_lock_match &&
824                     req->rq_ops->hpreq_lock_match(req, lock))
825                         ptlrpc_nrs_req_hp_move(req);
826         }
827         spin_unlock(&lock->l_export->exp_rpc_lock);
828         EXIT;
829 }
830
831 /**
832  * ->l_blocking_ast() method for server-side locks. This is invoked when newly
833  * enqueued server lock conflicts with given one.
834  *
835  * Sends blocking AST RPC to the client owning that lock; arms timeout timer
836  * to wait for client response.
837  */
838 int ldlm_server_blocking_ast(struct ldlm_lock *lock,
839                              struct ldlm_lock_desc *desc,
840                              void *data, int flag)
841 {
842         struct ldlm_cb_async_args *ca;
843         struct ldlm_cb_set_arg *arg = data;
844         struct ldlm_request    *body;
845         struct ptlrpc_request  *req;
846         int                     instant_cancel = 0;
847         int                     rc = 0;
848         ENTRY;
849
850         if (flag == LDLM_CB_CANCELING)
851                 /* Don't need to do anything here. */
852                 RETURN(0);
853
854         if (OBD_FAIL_PRECHECK(OBD_FAIL_LDLM_SRV_BL_AST)) {
855                 LDLM_DEBUG(lock, "dropping BL AST");
856                 RETURN(0);
857         }
858
859         LASSERT(lock);
860         LASSERT(data != NULL);
861         if (lock->l_export->exp_obd->obd_recovering != 0)
862                 LDLM_ERROR(lock, "BUG 6063: lock collide during recovery");
863
864         ldlm_lock_reorder_req(lock);
865
866         req = ptlrpc_request_alloc_pack(lock->l_export->exp_imp_reverse,
867                                         &RQF_LDLM_BL_CALLBACK,
868                                         LUSTRE_DLM_VERSION, LDLM_BL_CALLBACK);
869         if (req == NULL)
870                 RETURN(-ENOMEM);
871
872         CLASSERT(sizeof(*ca) <= sizeof(req->rq_async_args));
873         ca = ptlrpc_req_async_args(req);
874         ca->ca_set_arg = arg;
875         ca->ca_lock = lock;
876
877         req->rq_interpret_reply = ldlm_cb_interpret;
878
879         lock_res_and_lock(lock);
880         if (ldlm_is_destroyed(lock)) {
881                 /* What's the point? */
882                 unlock_res_and_lock(lock);
883                 ptlrpc_req_finished(req);
884                 RETURN(0);
885         }
886
887         if (lock->l_granted_mode != lock->l_req_mode) {
888                 /* this blocking AST will be communicated as part of the
889                  * completion AST instead */
890                 ldlm_add_blocked_lock(lock);
891                 ldlm_set_waited(lock);
892                 unlock_res_and_lock(lock);
893
894                 ptlrpc_req_finished(req);
895                 LDLM_DEBUG(lock, "lock not granted, not sending blocking AST");
896                 RETURN(0);
897         }
898
899         if (ldlm_is_cancel_on_block(lock))
900                 instant_cancel = 1;
901
902         body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
903         body->lock_handle[0] = lock->l_remote_handle;
904         body->lock_desc = *desc;
905         body->lock_flags |= ldlm_flags_to_wire(lock->l_flags & LDLM_FL_AST_MASK);
906
907         LDLM_DEBUG(lock, "server preparing blocking AST");
908
909         ptlrpc_request_set_replen(req);
910         ldlm_set_cbpending(lock);
911         if (instant_cancel) {
912                 unlock_res_and_lock(lock);
913                 ldlm_lock_cancel(lock);
914
915                 req->rq_no_resend = 1;
916         } else {
917                 LASSERT(lock->l_granted_mode == lock->l_req_mode);
918                 ldlm_add_waiting_lock(lock, ldlm_bl_timeout(lock));
919                 unlock_res_and_lock(lock);
920
921                 /* Do not resend after lock callback timeout */
922                 req->rq_delay_limit = ldlm_bl_timeout(lock);
923                 req->rq_resend_cb = ldlm_update_resend;
924         }
925
926         req->rq_send_state = LUSTRE_IMP_FULL;
927         /* ptlrpc_request_alloc_pack already set timeout */
928         if (AT_OFF)
929                 req->rq_timeout = ldlm_get_rq_timeout();
930
931         if (lock->l_export && lock->l_export->exp_nid_stats &&
932             lock->l_export->exp_nid_stats->nid_ldlm_stats)
933                 lprocfs_counter_incr(lock->l_export->exp_nid_stats->nid_ldlm_stats,
934                                      LDLM_BL_CALLBACK - LDLM_FIRST_OPC);
935
936         rc = ldlm_ast_fini(req, arg, lock, instant_cancel);
937
938         RETURN(rc);
939 }
940
941 /**
942  * ->l_completion_ast callback for a remote lock in server namespace.
943  *
944  *  Sends AST to the client notifying it of lock granting.  If initial
945  *  lock response was not sent yet, instead of sending another RPC, just
946  *  mark the lock as granted and client will understand
947  */
948 int ldlm_server_completion_ast(struct ldlm_lock *lock, __u64 flags, void *data)
949 {
950         struct ldlm_cb_set_arg *arg = data;
951         struct ldlm_request    *body;
952         struct ptlrpc_request  *req;
953         struct ldlm_cb_async_args *ca;
954         int                     instant_cancel = 0;
955         int                     rc = 0;
956         int                     lvb_len;
957         ENTRY;
958
959         LASSERT(lock != NULL);
960         LASSERT(data != NULL);
961
962         if (OBD_FAIL_PRECHECK(OBD_FAIL_LDLM_SRV_CP_AST)) {
963                 LDLM_DEBUG(lock, "dropping CP AST");
964                 RETURN(0);
965         }
966
967         req = ptlrpc_request_alloc(lock->l_export->exp_imp_reverse,
968                                     &RQF_LDLM_CP_CALLBACK);
969         if (req == NULL)
970                 RETURN(-ENOMEM);
971
972         /* server namespace, doesn't need lock */
973         lvb_len = ldlm_lvbo_size(lock);
974         /* LU-3124 & LU-2187: to not return layout in completion AST because
975          * it may deadlock for LU-2187, or client may not have enough space
976          * for large layout. The layout will be returned to client with an
977          * extra RPC to fetch xattr.lov */
978         if (ldlm_has_layout(lock))
979                 lvb_len = 0;
980
981         req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_CLIENT, lvb_len);
982         rc = ptlrpc_request_pack(req, LUSTRE_DLM_VERSION, LDLM_CP_CALLBACK);
983         if (rc) {
984                 ptlrpc_request_free(req);
985                 RETURN(rc);
986         }
987
988         CLASSERT(sizeof(*ca) <= sizeof(req->rq_async_args));
989         ca = ptlrpc_req_async_args(req);
990         ca->ca_set_arg = arg;
991         ca->ca_lock = lock;
992
993         req->rq_interpret_reply = ldlm_cb_interpret;
994         body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
995
996         body->lock_handle[0] = lock->l_remote_handle;
997         body->lock_flags = ldlm_flags_to_wire(flags);
998         ldlm_lock2desc(lock, &body->lock_desc);
999         if (lvb_len > 0) {
1000                 void *lvb = req_capsule_client_get(&req->rq_pill, &RMF_DLM_LVB);
1001                 const struct lu_env *env = NULL;
1002
1003                 if (req->rq_svc_thread)
1004                         env = req->rq_svc_thread->t_env;
1005
1006                 lvb_len = ldlm_lvbo_fill(env, lock, lvb, &lvb_len);
1007                 if (lvb_len < 0) {
1008                         /* We still need to send the RPC to wake up the blocked
1009                          * enqueue thread on the client.
1010                          *
1011                          * Consider old client, there is no better way to notify
1012                          * the failure, just zero-sized the LVB, then the client
1013                          * will fail out as "-EPROTO". */
1014                         req_capsule_shrink(&req->rq_pill, &RMF_DLM_LVB, 0,
1015                                            RCL_CLIENT);
1016                         instant_cancel = 1;
1017                 } else {
1018                         req_capsule_shrink(&req->rq_pill, &RMF_DLM_LVB, lvb_len,
1019                                            RCL_CLIENT);
1020                 }
1021         }
1022
1023         LDLM_DEBUG(lock, "server preparing completion AST");
1024
1025         ptlrpc_request_set_replen(req);
1026
1027         req->rq_send_state = LUSTRE_IMP_FULL;
1028         /* ptlrpc_request_pack already set timeout */
1029         if (AT_OFF)
1030                 req->rq_timeout = ldlm_get_rq_timeout();
1031
1032         /* We only send real blocking ASTs after the lock is granted */
1033         lock_res_and_lock(lock);
1034         if (ldlm_is_ast_sent(lock)) {
1035                 body->lock_flags |= ldlm_flags_to_wire(LDLM_FL_AST_SENT);
1036                 /* Copy AST flags like LDLM_FL_DISCARD_DATA. */
1037                 body->lock_flags |= ldlm_flags_to_wire(lock->l_flags &
1038                                                        LDLM_FL_AST_MASK);
1039
1040                 /* We might get here prior to ldlm_handle_enqueue setting
1041                  * LDLM_FL_CANCEL_ON_BLOCK flag. Then we will put this lock
1042                  * into waiting list, but this is safe and similar code in
1043                  * ldlm_handle_enqueue will call ldlm_lock_cancel() still,
1044                  * that would not only cancel the lock, but will also remove
1045                  * it from waiting list */
1046                 if (ldlm_is_cancel_on_block(lock)) {
1047                         unlock_res_and_lock(lock);
1048                         ldlm_lock_cancel(lock);
1049
1050                         instant_cancel = 1;
1051                         req->rq_no_resend = 1;
1052
1053                         lock_res_and_lock(lock);
1054                 } else {
1055                         /* start the lock-timeout clock */
1056                         ldlm_add_waiting_lock(lock, ldlm_bl_timeout(lock));
1057                         /* Do not resend after lock callback timeout */
1058                         req->rq_delay_limit = ldlm_bl_timeout(lock);
1059                         req->rq_resend_cb = ldlm_update_resend;
1060                 }
1061         }
1062         unlock_res_and_lock(lock);
1063
1064         if (lock->l_export && lock->l_export->exp_nid_stats &&
1065             lock->l_export->exp_nid_stats->nid_ldlm_stats)
1066                 lprocfs_counter_incr(lock->l_export->exp_nid_stats->nid_ldlm_stats,
1067                                      LDLM_CP_CALLBACK - LDLM_FIRST_OPC);
1068
1069         rc = ldlm_ast_fini(req, arg, lock, instant_cancel);
1070
1071         RETURN(lvb_len < 0 ? lvb_len : rc);
1072 }
1073
1074 /**
1075  * Server side ->l_glimpse_ast handler for client locks.
1076  *
1077  * Sends glimpse AST to the client and waits for reply. Then updates
1078  * lvbo with the result.
1079  */
1080 int ldlm_server_glimpse_ast(struct ldlm_lock *lock, void *data)
1081 {
1082         struct ldlm_cb_set_arg          *arg = data;
1083         struct ldlm_request             *body;
1084         struct ptlrpc_request           *req;
1085         struct ldlm_cb_async_args       *ca;
1086         int                              rc;
1087         struct req_format               *req_fmt;
1088         ENTRY;
1089
1090         LASSERT(lock != NULL);
1091
1092         if (arg->gl_desc != NULL)
1093                 /* There is a glimpse descriptor to pack */
1094                 req_fmt = &RQF_LDLM_GL_CALLBACK_DESC;
1095         else
1096                 req_fmt = &RQF_LDLM_GL_CALLBACK;
1097
1098         req = ptlrpc_request_alloc_pack(lock->l_export->exp_imp_reverse,
1099                                         req_fmt, LUSTRE_DLM_VERSION,
1100                                         LDLM_GL_CALLBACK);
1101
1102         if (req == NULL)
1103                 RETURN(-ENOMEM);
1104
1105         if (arg->gl_desc != NULL) {
1106                 /* copy the GL descriptor */
1107                 union ldlm_gl_desc      *desc;
1108                 desc = req_capsule_client_get(&req->rq_pill, &RMF_DLM_GL_DESC);
1109                 *desc = *arg->gl_desc;
1110         }
1111
1112         body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1113         body->lock_handle[0] = lock->l_remote_handle;
1114         ldlm_lock2desc(lock, &body->lock_desc);
1115
1116         CLASSERT(sizeof(*ca) <= sizeof(req->rq_async_args));
1117         ca = ptlrpc_req_async_args(req);
1118         ca->ca_set_arg = arg;
1119         ca->ca_lock = lock;
1120
1121         /* server namespace, doesn't need lock */
1122         req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
1123                              ldlm_lvbo_size(lock));
1124         ptlrpc_request_set_replen(req);
1125
1126         req->rq_send_state = LUSTRE_IMP_FULL;
1127         /* ptlrpc_request_alloc_pack already set timeout */
1128         if (AT_OFF)
1129                 req->rq_timeout = ldlm_get_rq_timeout();
1130
1131         req->rq_interpret_reply = ldlm_cb_interpret;
1132
1133         if (lock->l_export && lock->l_export->exp_nid_stats &&
1134             lock->l_export->exp_nid_stats->nid_ldlm_stats)
1135                 lprocfs_counter_incr(lock->l_export->exp_nid_stats->nid_ldlm_stats,
1136                                      LDLM_GL_CALLBACK - LDLM_FIRST_OPC);
1137
1138         rc = ldlm_ast_fini(req, arg, lock, 0);
1139
1140         RETURN(rc);
1141 }
1142 EXPORT_SYMBOL(ldlm_server_glimpse_ast);
1143
1144 int ldlm_glimpse_locks(struct ldlm_resource *res,
1145                        struct list_head *gl_work_list)
1146 {
1147         int     rc;
1148         ENTRY;
1149
1150         rc = ldlm_run_ast_work(ldlm_res_to_ns(res), gl_work_list,
1151                                LDLM_WORK_GL_AST);
1152         if (rc == -ERESTART)
1153                 ldlm_reprocess_all(res);
1154
1155         RETURN(rc);
1156 }
1157 EXPORT_SYMBOL(ldlm_glimpse_locks);
1158
1159 /* return LDLM lock associated with a lock callback request */
1160 struct ldlm_lock *ldlm_request_lock(struct ptlrpc_request *req)
1161 {
1162         struct ldlm_cb_async_args       *ca;
1163         struct ldlm_lock                *lock;
1164         ENTRY;
1165
1166         ca = ptlrpc_req_async_args(req);
1167         lock = ca->ca_lock;
1168         if (lock == NULL)
1169                 RETURN(ERR_PTR(-EFAULT));
1170
1171         RETURN(lock);
1172 }
1173 EXPORT_SYMBOL(ldlm_request_lock);
1174
1175 /**
1176  * Main server-side entry point into LDLM for enqueue. This is called by ptlrpc
1177  * service threads to carry out client lock enqueueing requests.
1178  */
1179 int ldlm_handle_enqueue0(struct ldlm_namespace *ns,
1180                          struct ptlrpc_request *req,
1181                          const struct ldlm_request *dlm_req,
1182                          const struct ldlm_callback_suite *cbs)
1183 {
1184         struct ldlm_reply *dlm_rep;
1185         __u64 flags;
1186         enum ldlm_error err = ELDLM_OK;
1187         struct ldlm_lock *lock = NULL;
1188         void *cookie = NULL;
1189         int rc = 0;
1190         struct ldlm_resource *res = NULL;
1191         const struct lu_env *env = req->rq_svc_thread->t_env;
1192         ENTRY;
1193
1194         LDLM_DEBUG_NOLOCK("server-side enqueue handler START");
1195
1196         ldlm_request_cancel(req, dlm_req, LDLM_ENQUEUE_CANCEL_OFF, LATF_SKIP);
1197         flags = ldlm_flags_from_wire(dlm_req->lock_flags);
1198
1199         LASSERT(req->rq_export);
1200
1201         /* for intent enqueue the stat will be updated inside intent policy */
1202         if (ptlrpc_req2svc(req)->srv_stats != NULL &&
1203             !(dlm_req->lock_flags & LDLM_FL_HAS_INTENT))
1204                 ldlm_svc_get_eopc(dlm_req, ptlrpc_req2svc(req)->srv_stats);
1205
1206         if (req->rq_export && req->rq_export->exp_nid_stats &&
1207             req->rq_export->exp_nid_stats->nid_ldlm_stats)
1208                 lprocfs_counter_incr(req->rq_export->exp_nid_stats->nid_ldlm_stats,
1209                                      LDLM_ENQUEUE - LDLM_FIRST_OPC);
1210
1211         if (unlikely(dlm_req->lock_desc.l_resource.lr_type < LDLM_MIN_TYPE ||
1212                      dlm_req->lock_desc.l_resource.lr_type >= LDLM_MAX_TYPE)) {
1213                 DEBUG_REQ(D_ERROR, req, "invalid lock request type %d",
1214                           dlm_req->lock_desc.l_resource.lr_type);
1215                 GOTO(out, rc = -EFAULT);
1216         }
1217
1218         if (unlikely(dlm_req->lock_desc.l_req_mode <= LCK_MINMODE ||
1219                      dlm_req->lock_desc.l_req_mode >= LCK_MAXMODE ||
1220                      dlm_req->lock_desc.l_req_mode &
1221                      (dlm_req->lock_desc.l_req_mode-1))) {
1222                 DEBUG_REQ(D_ERROR, req, "invalid lock request mode %d",
1223                           dlm_req->lock_desc.l_req_mode);
1224                 GOTO(out, rc = -EFAULT);
1225         }
1226
1227         if (unlikely((flags & LDLM_FL_REPLAY) ||
1228                      (lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT))) {
1229                 /* Find an existing lock in the per-export lock hash */
1230                 /* In the function below, .hs_keycmp resolves to
1231                  * ldlm_export_lock_keycmp() */
1232                 /* coverity[overrun-buffer-val] */
1233                 lock = cfs_hash_lookup(req->rq_export->exp_lock_hash,
1234                                        (void *)&dlm_req->lock_handle[0]);
1235                 if (lock != NULL) {
1236                         DEBUG_REQ(D_DLMTRACE, req, "found existing lock cookie %#llx",
1237                                   lock->l_handle.h_cookie);
1238                         flags |= LDLM_FL_RESENT;
1239                         GOTO(existing_lock, rc = 0);
1240                 }
1241         } else {
1242                 if (ldlm_reclaim_full()) {
1243                         DEBUG_REQ(D_DLMTRACE, req, "Too many granted locks, "
1244                                   "reject current enqueue request and let the "
1245                                   "client retry later.\n");
1246                         GOTO(out, rc = -EINPROGRESS);
1247                 }
1248         }
1249
1250         /* The lock's callback data might be set in the policy function */
1251         lock = ldlm_lock_create(ns, &dlm_req->lock_desc.l_resource.lr_name,
1252                                 dlm_req->lock_desc.l_resource.lr_type,
1253                                 dlm_req->lock_desc.l_req_mode,
1254                                 cbs, NULL, 0, LVB_T_NONE);
1255         if (IS_ERR(lock)) {
1256                 rc = PTR_ERR(lock);
1257                 lock = NULL;
1258                 GOTO(out, rc);
1259         }
1260
1261         lock->l_remote_handle = dlm_req->lock_handle[0];
1262         LDLM_DEBUG(lock, "server-side enqueue handler, new lock created");
1263
1264         /* Initialize resource lvb but not for a lock being replayed since
1265          * Client already got lvb sent in this case.
1266          * This must occur early since some policy methods assume resource
1267          * lvb is available (lr_lvb_data != NULL).
1268          */
1269         res = lock->l_resource;
1270         if (!(flags & LDLM_FL_REPLAY)) {
1271                 /* non-replayed lock, delayed lvb init may need to be done */
1272                 rc = ldlm_lvbo_init(env, res);
1273                 if (rc < 0) {
1274                         LDLM_DEBUG(lock, "delayed lvb init failed (rc %d)", rc);
1275                         GOTO(out, rc);
1276                 }
1277         }
1278
1279         OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_ENQUEUE_BLOCKED, obd_timeout * 2);
1280         /* Don't enqueue a lock onto the export if it is been disonnected
1281          * due to eviction (bug 3822) or server umount (bug 24324).
1282          * Cancel it now instead. */
1283         if (req->rq_export->exp_disconnected) {
1284                 LDLM_ERROR(lock, "lock on disconnected export %p",
1285                            req->rq_export);
1286                 GOTO(out, rc = -ENOTCONN);
1287         }
1288
1289         lock->l_export = class_export_lock_get(req->rq_export, lock);
1290         if (lock->l_export->exp_lock_hash)
1291                 cfs_hash_add(lock->l_export->exp_lock_hash,
1292                              &lock->l_remote_handle,
1293                              &lock->l_exp_hash);
1294
1295         /* Inherit the enqueue flags before the operation, because we do not
1296          * keep the res lock on return and next operations (BL AST) may proceed
1297          * without them. */
1298         lock->l_flags |= ldlm_flags_from_wire(dlm_req->lock_flags &
1299                                               LDLM_FL_INHERIT_MASK);
1300
1301         ldlm_convert_policy_to_local(req->rq_export,
1302                                      dlm_req->lock_desc.l_resource.lr_type,
1303                                      &dlm_req->lock_desc.l_policy_data,
1304                                      &lock->l_policy_data);
1305         if (dlm_req->lock_desc.l_resource.lr_type == LDLM_EXTENT)
1306                 lock->l_req_extent = lock->l_policy_data.l_extent;
1307
1308 existing_lock:
1309         if (flags & LDLM_FL_HAS_INTENT) {
1310                 /* In this case, the reply buffer is allocated deep in
1311                  * local_lock_enqueue by the policy function. */
1312                 cookie = req;
1313         } else {
1314                 /* based on the assumption that lvb size never changes during
1315                  * resource life time otherwise it need resource->lr_lock's
1316                  * protection */
1317                 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB,
1318                                      RCL_SERVER, ldlm_lvbo_size(lock));
1319
1320                 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_ENQUEUE_EXTENT_ERR))
1321                         GOTO(out, rc = -ENOMEM);
1322
1323                 rc = req_capsule_server_pack(&req->rq_pill);
1324                 if (rc)
1325                         GOTO(out, rc);
1326         }
1327
1328         err = ldlm_lock_enqueue(env, ns, &lock, cookie, &flags);
1329         if (err) {
1330                 if ((int)err < 0)
1331                         rc = (int)err;
1332                 GOTO(out, err);
1333         }
1334
1335         dlm_rep = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
1336
1337         ldlm_lock2desc(lock, &dlm_rep->lock_desc);
1338         ldlm_lock2handle(lock, &dlm_rep->lock_handle);
1339
1340         if (lock && lock->l_resource->lr_type == LDLM_EXTENT)
1341                 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_BL_EVICT, 6);
1342
1343         /* We never send a blocking AST until the lock is granted, but
1344          * we can tell it right now */
1345         lock_res_and_lock(lock);
1346
1347         /* Now take into account flags to be inherited from original lock
1348            request both in reply to client and in our own lock flags. */
1349         dlm_rep->lock_flags = ldlm_flags_to_wire(flags);
1350         lock->l_flags |= flags & LDLM_FL_INHERIT_MASK;
1351
1352         /* Don't move a pending lock onto the export if it has already been
1353          * disconnected due to eviction (bug 5683) or server umount (bug 24324).
1354          * Cancel it now instead. */
1355         if (unlikely(req->rq_export->exp_disconnected ||
1356                      OBD_FAIL_CHECK(OBD_FAIL_LDLM_ENQUEUE_OLD_EXPORT))) {
1357                 LDLM_ERROR(lock, "lock on destroyed export %p", req->rq_export);
1358                 rc = -ENOTCONN;
1359         } else if (ldlm_is_ast_sent(lock)) {
1360                 /* fill lock desc for possible lock convert */
1361                 if (lock->l_blocking_lock &&
1362                     lock->l_resource->lr_type == LDLM_IBITS) {
1363                         struct ldlm_lock *bl_lock = lock->l_blocking_lock;
1364                         struct ldlm_lock_desc *rep_desc = &dlm_rep->lock_desc;
1365
1366                         LDLM_DEBUG(lock,
1367                                    "save blocking bits %llx in granted lock",
1368                                    bl_lock->l_policy_data.l_inodebits.bits);
1369                         /* If lock is blocked then save blocking ibits
1370                          * in returned lock policy for the possible lock
1371                          * convert on a client.
1372                          */
1373                         rep_desc->l_policy_data.l_inodebits.cancel_bits =
1374                                 bl_lock->l_policy_data.l_inodebits.bits;
1375                 }
1376                 dlm_rep->lock_flags |= ldlm_flags_to_wire(LDLM_FL_AST_SENT);
1377                 if (lock->l_granted_mode == lock->l_req_mode) {
1378                         /*
1379                          * Only cancel lock if it was granted, because it would
1380                          * be destroyed immediately and would never be granted
1381                          * in the future, causing timeouts on client.  Not
1382                          * granted lock will be cancelled immediately after
1383                          * sending completion AST.
1384                          */
1385                         if (dlm_rep->lock_flags & LDLM_FL_CANCEL_ON_BLOCK) {
1386                                 unlock_res_and_lock(lock);
1387                                 ldlm_lock_cancel(lock);
1388                                 lock_res_and_lock(lock);
1389                         } else {
1390                                 ldlm_add_waiting_lock(lock,
1391                                                       ldlm_bl_timeout(lock));
1392                         }
1393                 }
1394         }
1395         unlock_res_and_lock(lock);
1396
1397         EXIT;
1398  out:
1399         req->rq_status = rc ?: err; /* return either error - bug 11190 */
1400         if (!req->rq_packed_final) {
1401                 err = lustre_pack_reply(req, 1, NULL, NULL);
1402                 if (rc == 0)
1403                         rc = err;
1404         }
1405
1406         /* The LOCK_CHANGED code in ldlm_lock_enqueue depends on this
1407          * ldlm_reprocess_all.  If this moves, revisit that code. -phil */
1408         if (lock != NULL) {
1409                 LDLM_DEBUG(lock, "server-side enqueue handler, sending reply"
1410                            "(err=%d, rc=%d)", err, rc);
1411
1412                 if (rc == 0 &&
1413                     req_capsule_has_field(&req->rq_pill, &RMF_DLM_LVB,
1414                                           RCL_SERVER) &&
1415                     ldlm_lvbo_size(lock) > 0) {
1416                         void *buf;
1417                         int buflen;
1418
1419 retry:
1420                         buf = req_capsule_server_get(&req->rq_pill,
1421                                                      &RMF_DLM_LVB);
1422                         LASSERTF(buf != NULL, "req %p, lock %p\n", req, lock);
1423                         buflen = req_capsule_get_size(&req->rq_pill,
1424                                         &RMF_DLM_LVB, RCL_SERVER);
1425                         /* non-replayed lock, delayed lvb init may
1426                          * need to be occur now
1427                          */
1428                         if ((buflen > 0) && !(flags & LDLM_FL_REPLAY)) {
1429                                 int rc2;
1430
1431                                 rc2 = ldlm_lvbo_fill(env, lock, buf, &buflen);
1432                                 if (rc2 >= 0) {
1433                                         req_capsule_shrink(&req->rq_pill,
1434                                                            &RMF_DLM_LVB,
1435                                                            rc2, RCL_SERVER);
1436                                 } else if (rc2 == -ERANGE) {
1437                                         rc2 = req_capsule_server_grow(
1438                                                         &req->rq_pill,
1439                                                         &RMF_DLM_LVB, buflen);
1440                                         if (!rc2) {
1441                                                 goto retry;
1442                                         } else {
1443                                                 /* if we can't grow the buffer,
1444                                                  * it's ok to return empty lvb
1445                                                  * to client.
1446                                                  */
1447                                                 req_capsule_shrink(
1448                                                         &req->rq_pill,
1449                                                         &RMF_DLM_LVB, 0,
1450                                                         RCL_SERVER);
1451                                         }
1452                                 } else {
1453                                         rc = rc2;
1454                                 }
1455                         } else if (flags & LDLM_FL_REPLAY) {
1456                                 /* no LVB resend upon replay */
1457                                 if (buflen > 0)
1458                                         req_capsule_shrink(&req->rq_pill,
1459                                                            &RMF_DLM_LVB,
1460                                                            0, RCL_SERVER);
1461                                 else
1462                                         rc = buflen;
1463                         } else {
1464                                 rc = buflen;
1465                         }
1466                 }
1467
1468                 if (rc != 0 && !(flags & LDLM_FL_RESENT)) {
1469                         if (lock->l_export) {
1470                                 ldlm_lock_cancel(lock);
1471                         } else {
1472                                 lock_res_and_lock(lock);
1473                                 ldlm_resource_unlink_lock(lock);
1474                                 ldlm_lock_destroy_nolock(lock);
1475                                 unlock_res_and_lock(lock);
1476
1477                         }
1478                 }
1479
1480                 if (!err && !ldlm_is_cbpending(lock) &&
1481                     dlm_req->lock_desc.l_resource.lr_type != LDLM_FLOCK)
1482                         ldlm_reprocess_all(lock->l_resource);
1483
1484                 LDLM_LOCK_RELEASE(lock);
1485         }
1486
1487         LDLM_DEBUG_NOLOCK("server-side enqueue handler END (lock %p, rc %d)",
1488                           lock, rc);
1489
1490         return rc;
1491 }
1492
1493 /**
1494  * Main LDLM entry point for server code to process lock conversion requests.
1495  */
1496 int ldlm_handle_convert0(struct ptlrpc_request *req,
1497                          const struct ldlm_request *dlm_req)
1498 {
1499         struct obd_export *exp = req->rq_export;
1500         struct ldlm_reply *dlm_rep;
1501         struct ldlm_lock *lock;
1502         int rc;
1503
1504         ENTRY;
1505
1506         if (exp && exp->exp_nid_stats && exp->exp_nid_stats->nid_ldlm_stats)
1507                 lprocfs_counter_incr(exp->exp_nid_stats->nid_ldlm_stats,
1508                                      LDLM_CONVERT - LDLM_FIRST_OPC);
1509
1510         rc = req_capsule_server_pack(&req->rq_pill);
1511         if (rc)
1512                 RETURN(rc);
1513
1514         dlm_rep = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
1515         dlm_rep->lock_flags = dlm_req->lock_flags;
1516
1517         lock = ldlm_handle2lock(&dlm_req->lock_handle[0]);
1518         if (lock) {
1519                 __u64 bits;
1520                 __u64 new;
1521
1522                 bits = lock->l_policy_data.l_inodebits.bits;
1523                 new = dlm_req->lock_desc.l_policy_data.l_inodebits.bits;
1524                 LDLM_DEBUG(lock, "server-side convert handler START");
1525
1526                 if (ldlm_is_cancel(lock)) {
1527                         LDLM_ERROR(lock, "convert on canceled lock!");
1528                         rc = ELDLM_NO_LOCK_DATA;
1529                 } else if (dlm_req->lock_desc.l_req_mode !=
1530                            lock->l_granted_mode) {
1531                         LDLM_ERROR(lock, "lock mode differs!");
1532                         rc = ELDLM_NO_LOCK_DATA;
1533                 } else if (bits == new) {
1534                         /* This can be valid situation if CONVERT RPCs are
1535                          * re-ordered. Just finish silently */
1536                         LDLM_DEBUG(lock, "lock is converted already!");
1537                         rc = ELDLM_OK;
1538                 } else {
1539                         lock_res_and_lock(lock);
1540                         if (ldlm_is_waited(lock))
1541                                 ldlm_del_waiting_lock(lock);
1542
1543                         ldlm_clear_cbpending(lock);
1544                         lock->l_policy_data.l_inodebits.cancel_bits = 0;
1545                         ldlm_inodebits_drop(lock, bits & ~new);
1546                         /* if lock is in a bl_ast list, remove it from the list
1547                          * here before reprocessing.
1548                          */
1549                         if (!list_empty(&lock->l_bl_ast)) {
1550                                 ldlm_discard_bl_lock(lock);
1551                         } else {
1552                                 /* in this case lock was taken from bl_ast list
1553                                  * already by ldlm_work_bl_ast_lock() and lock
1554                                  * must clear only some remaining states.
1555                                  */
1556                                 ldlm_clear_ast_sent(lock);
1557                                 lock->l_bl_ast_run = 0;
1558                                 ldlm_clear_blocking_lock(lock);
1559                         }
1560                         unlock_res_and_lock(lock);
1561
1562                         ldlm_reprocess_all(lock->l_resource);
1563                         rc = ELDLM_OK;
1564                 }
1565
1566                 if (rc == ELDLM_OK) {
1567                         dlm_rep->lock_handle = lock->l_remote_handle;
1568                         ldlm_ibits_policy_local_to_wire(&lock->l_policy_data,
1569                                         &dlm_rep->lock_desc.l_policy_data);
1570                 }
1571
1572                 LDLM_DEBUG(lock, "server-side convert handler END, rc = %d",
1573                            rc);
1574                 LDLM_LOCK_PUT(lock);
1575         } else {
1576                 rc = ELDLM_NO_LOCK_DATA;
1577                 LDLM_DEBUG_NOLOCK("server-side convert handler END, rc = %d",
1578                                   rc);
1579         }
1580
1581         req->rq_status = rc;
1582
1583         RETURN(0);
1584 }
1585
1586 /**
1587  * Cancel all the locks whose handles are packed into ldlm_request
1588  *
1589  * Called by server code expecting such combined cancel activity
1590  * requests.
1591  */
1592 int ldlm_request_cancel(struct ptlrpc_request *req,
1593                         const struct ldlm_request *dlm_req,
1594                         int first, enum lustre_at_flags flags)
1595 {
1596         const struct lu_env *env = req->rq_svc_thread->t_env;
1597         struct ldlm_resource *res, *pres = NULL;
1598         struct ldlm_lock *lock;
1599         int i, count, done = 0;
1600         ENTRY;
1601
1602         count = dlm_req->lock_count ? dlm_req->lock_count : 1;
1603         if (first >= count)
1604                 RETURN(0);
1605
1606         if (count == 1 && dlm_req->lock_handle[0].cookie == 0)
1607                 RETURN(0);
1608
1609         /* There is no lock on the server at the replay time,
1610          * skip lock cancelling to make replay tests to pass. */
1611         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
1612                 RETURN(0);
1613
1614         LDLM_DEBUG_NOLOCK("server-side cancel handler START: %d locks, "
1615                           "starting at %d", count, first);
1616
1617         for (i = first; i < count; i++) {
1618                 lock = ldlm_handle2lock(&dlm_req->lock_handle[i]);
1619                 if (!lock) {
1620                         LDLM_DEBUG_NOLOCK("server-side cancel handler stale "
1621                                           "lock (cookie %llu)",
1622                                           dlm_req->lock_handle[i].cookie);
1623                         continue;
1624                 }
1625
1626                 res = lock->l_resource;
1627                 done++;
1628
1629                 /* This code is an optimization to only attempt lock
1630                  * granting on the resource (that could be CPU-expensive)
1631                  * after we are done cancelling lock in that resource. */
1632                 if (res != pres) {
1633                         if (pres != NULL) {
1634                                 ldlm_reprocess_all(pres);
1635                                 LDLM_RESOURCE_DELREF(pres);
1636                                 ldlm_resource_putref(pres);
1637                         }
1638                         if (res != NULL) {
1639                                 ldlm_resource_getref(res);
1640                                 LDLM_RESOURCE_ADDREF(res);
1641
1642                                 if (!ldlm_is_discard_data(lock))
1643                                         ldlm_lvbo_update(env, res, lock,
1644                                                          NULL, 1);
1645                         }
1646                         pres = res;
1647                 }
1648
1649                 if ((flags & LATF_STATS) && ldlm_is_ast_sent(lock) &&
1650                     lock->l_blast_sent != 0) {
1651                         time64_t delay = ktime_get_real_seconds() -
1652                                          lock->l_blast_sent;
1653                         LDLM_DEBUG(lock, "server cancels blocked lock after %llds",
1654                                    (s64)delay);
1655                         at_measured(&lock->l_export->exp_bl_lock_at, delay);
1656                 }
1657                 ldlm_lock_cancel(lock);
1658                 LDLM_LOCK_PUT(lock);
1659         }
1660         if (pres != NULL) {
1661                 ldlm_reprocess_all(pres);
1662                 LDLM_RESOURCE_DELREF(pres);
1663                 ldlm_resource_putref(pres);
1664         }
1665         LDLM_DEBUG_NOLOCK("server-side cancel handler END");
1666         RETURN(done);
1667 }
1668 EXPORT_SYMBOL(ldlm_request_cancel);
1669
1670 /**
1671  * Main LDLM entry point for server code to cancel locks.
1672  *
1673  * Typically gets called from service handler on LDLM_CANCEL opc.
1674  */
1675 int ldlm_handle_cancel(struct ptlrpc_request *req)
1676 {
1677         struct ldlm_request *dlm_req;
1678         int rc;
1679         ENTRY;
1680
1681         dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1682         if (dlm_req == NULL) {
1683                 CDEBUG(D_INFO, "bad request buffer for cancel\n");
1684                 RETURN(-EFAULT);
1685         }
1686
1687         if (req->rq_export && req->rq_export->exp_nid_stats &&
1688             req->rq_export->exp_nid_stats->nid_ldlm_stats)
1689                 lprocfs_counter_incr(req->rq_export->exp_nid_stats->nid_ldlm_stats,
1690                                      LDLM_CANCEL - LDLM_FIRST_OPC);
1691
1692         rc = req_capsule_server_pack(&req->rq_pill);
1693         if (rc)
1694                 RETURN(rc);
1695
1696         if (!ldlm_request_cancel(req, dlm_req, 0, LATF_STATS))
1697                 req->rq_status = LUSTRE_ESTALE;
1698
1699         RETURN(ptlrpc_reply(req));
1700 }
1701 #endif /* HAVE_SERVER_SUPPORT */
1702
1703 /**
1704  * Callback handler for receiving incoming blocking ASTs.
1705  *
1706  * This can only happen on client side.
1707  */
1708 void ldlm_handle_bl_callback(struct ldlm_namespace *ns,
1709                              struct ldlm_lock_desc *ld, struct ldlm_lock *lock)
1710 {
1711         int do_ast;
1712         ENTRY;
1713
1714         LDLM_DEBUG(lock, "client blocking AST callback handler");
1715
1716         lock_res_and_lock(lock);
1717
1718         /* set bits to cancel for this lock for possible lock convert */
1719         if (ns_is_client(ns) && (lock->l_resource->lr_type == LDLM_IBITS)) {
1720                 /* Lock description contains policy of blocking lock,
1721                  * and its cancel_bits is used to pass conflicting bits.
1722                  * NOTE: ld can be NULL or can be not NULL but zeroed if
1723                  * passed from ldlm_bl_thread_blwi(), check below used bits
1724                  * in ld to make sure it is valid description.
1725                  */
1726                 if (ld && ld->l_policy_data.l_inodebits.bits)
1727                         lock->l_policy_data.l_inodebits.cancel_bits =
1728                                 ld->l_policy_data.l_inodebits.cancel_bits;
1729                 /* if there is no valid ld and lock is cbpending already
1730                  * then cancel_bits should be kept, otherwise it is zeroed.
1731                  */
1732                 else if (!ldlm_is_cbpending(lock))
1733                         lock->l_policy_data.l_inodebits.cancel_bits = 0;
1734         }
1735         ldlm_set_cbpending(lock);
1736
1737         if (ldlm_is_cancel_on_block(lock))
1738                 ldlm_set_cancel(lock);
1739
1740         do_ast = (!lock->l_readers && !lock->l_writers);
1741         unlock_res_and_lock(lock);
1742
1743         if (do_ast) {
1744                 CDEBUG(D_DLMTRACE, "Lock %p already unused, calling callback (%p)\n",
1745                        lock, lock->l_blocking_ast);
1746                 if (lock->l_blocking_ast != NULL)
1747                         lock->l_blocking_ast(lock, ld, lock->l_ast_data,
1748                                              LDLM_CB_BLOCKING);
1749         } else {
1750                 CDEBUG(D_DLMTRACE, "Lock %p is referenced, will be cancelled later\n",
1751                        lock);
1752         }
1753
1754         LDLM_DEBUG(lock, "client blocking callback handler END");
1755         LDLM_LOCK_RELEASE(lock);
1756         EXIT;
1757 }
1758
1759 /**
1760  * Callback handler for receiving incoming completion ASTs.
1761  *
1762  * This only can happen on client side.
1763  */
1764 static void ldlm_handle_cp_callback(struct ptlrpc_request *req,
1765                                     struct ldlm_namespace *ns,
1766                                     struct ldlm_request *dlm_req,
1767                                     struct ldlm_lock *lock)
1768 {
1769         struct list_head ast_list;
1770         int lvb_len;
1771         int rc = 0;
1772         ENTRY;
1773
1774         LDLM_DEBUG(lock, "client completion callback handler START");
1775
1776         INIT_LIST_HEAD(&ast_list);
1777         if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_BL_CB_RACE)) {
1778                 long to = cfs_time_seconds(1);
1779
1780                 while (to > 0) {
1781                         set_current_state(TASK_INTERRUPTIBLE);
1782                         schedule_timeout(to);
1783                         if (lock->l_granted_mode == lock->l_req_mode ||
1784                             ldlm_is_destroyed(lock))
1785                                 break;
1786                 }
1787         }
1788
1789         lvb_len = req_capsule_get_size(&req->rq_pill, &RMF_DLM_LVB, RCL_CLIENT);
1790         if (lvb_len < 0) {
1791                 LDLM_ERROR(lock, "Fail to get lvb_len, rc = %d", lvb_len);
1792                 GOTO(out, rc = lvb_len);
1793         } else if (lvb_len > 0) {
1794                 if (lock->l_lvb_len > 0) {
1795                         /* for extent lock, lvb contains ost_lvb{}. */
1796                         LASSERT(lock->l_lvb_data != NULL);
1797
1798                         if (unlikely(lock->l_lvb_len < lvb_len)) {
1799                                 LDLM_ERROR(lock, "Replied LVB is larger than "
1800                                            "expectation, expected = %d, "
1801                                            "replied = %d",
1802                                            lock->l_lvb_len, lvb_len);
1803                                 GOTO(out, rc = -EINVAL);
1804                         }
1805                 }
1806         }
1807
1808         lock_res_and_lock(lock);
1809
1810         if (!ldlm_res_eq(&dlm_req->lock_desc.l_resource.lr_name,
1811                          &lock->l_resource->lr_name)) {
1812                 ldlm_resource_unlink_lock(lock);
1813                 unlock_res_and_lock(lock);
1814                 rc = ldlm_lock_change_resource(ns, lock,
1815                                 &dlm_req->lock_desc.l_resource.lr_name);
1816                 if (rc < 0) {
1817                         LDLM_ERROR(lock, "Failed to allocate resource");
1818                         GOTO(out, rc);
1819                 }
1820                 LDLM_DEBUG(lock, "completion AST, new resource");
1821                 lock_res_and_lock(lock);
1822         }
1823
1824         if (ldlm_is_destroyed(lock) ||
1825             lock->l_granted_mode == lock->l_req_mode) {
1826                 /* bug 11300: the lock has already been granted */
1827                 unlock_res_and_lock(lock);
1828                 LDLM_DEBUG(lock, "Double grant race happened");
1829                 GOTO(out, rc = 0);
1830         }
1831
1832         /* If we receive the completion AST before the actual enqueue returned,
1833          * then we might need to switch lock modes, resources, or extents. */
1834         if (dlm_req->lock_desc.l_granted_mode != lock->l_req_mode) {
1835                 lock->l_req_mode = dlm_req->lock_desc.l_granted_mode;
1836                 LDLM_DEBUG(lock, "completion AST, new lock mode");
1837         }
1838
1839         if (lock->l_resource->lr_type != LDLM_PLAIN) {
1840                 ldlm_convert_policy_to_local(req->rq_export,
1841                                           dlm_req->lock_desc.l_resource.lr_type,
1842                                           &dlm_req->lock_desc.l_policy_data,
1843                                           &lock->l_policy_data);
1844                 LDLM_DEBUG(lock, "completion AST, new policy data");
1845         }
1846
1847         ldlm_resource_unlink_lock(lock);
1848
1849         if (dlm_req->lock_flags & LDLM_FL_AST_SENT) {
1850                 /* BL_AST locks are not needed in LRU.
1851                  * Let ldlm_cancel_lru() be fast. */
1852                 ldlm_lock_remove_from_lru(lock);
1853                 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_BL_AST;
1854                 LDLM_DEBUG(lock, "completion AST includes blocking AST");
1855         }
1856
1857         if (lock->l_lvb_len > 0) {
1858                 rc = ldlm_fill_lvb(lock, &req->rq_pill, RCL_CLIENT,
1859                                    lock->l_lvb_data, lvb_len);
1860                 if (rc < 0) {
1861                         unlock_res_and_lock(lock);
1862                         GOTO(out, rc);
1863                 }
1864         }
1865
1866         ldlm_grant_lock(lock, &ast_list);
1867         unlock_res_and_lock(lock);
1868
1869         LDLM_DEBUG(lock, "callback handler finished, about to run_ast_work");
1870
1871         /* Let Enqueue to call osc_lock_upcall() and initialize
1872          * l_ast_data */
1873         OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_CP_ENQ_RACE, 2);
1874
1875         ldlm_run_ast_work(ns, &ast_list, LDLM_WORK_CP_AST);
1876
1877         LDLM_DEBUG_NOLOCK("client completion callback handler END (lock %p)",
1878                           lock);
1879         GOTO(out, rc);
1880
1881 out:
1882         if (rc < 0) {
1883                 lock_res_and_lock(lock);
1884                 ldlm_set_failed(lock);
1885                 unlock_res_and_lock(lock);
1886                 wake_up(&lock->l_waitq);
1887         }
1888         LDLM_LOCK_RELEASE(lock);
1889 }
1890
1891 /**
1892  * Callback handler for receiving incoming glimpse ASTs.
1893  *
1894  * This only can happen on client side.  After handling the glimpse AST
1895  * we also consider dropping the lock here if it is unused locally for a
1896  * long time.
1897  */
1898 static void ldlm_handle_gl_callback(struct ptlrpc_request *req,
1899                                     struct ldlm_namespace *ns,
1900                                     struct ldlm_request *dlm_req,
1901                                     struct ldlm_lock *lock)
1902 {
1903         int rc = -ENOSYS;
1904         ENTRY;
1905
1906         LDLM_DEBUG(lock, "client glimpse AST callback handler");
1907
1908         if (lock->l_glimpse_ast != NULL)
1909                 rc = lock->l_glimpse_ast(lock, req);
1910
1911         if (req->rq_repmsg != NULL) {
1912                 ptlrpc_reply(req);
1913         } else {
1914                 req->rq_status = rc;
1915                 ptlrpc_error(req);
1916         }
1917
1918         lock_res_and_lock(lock);
1919         if (lock->l_granted_mode == LCK_PW &&
1920             !lock->l_readers && !lock->l_writers &&
1921             ktime_after(ktime_get(),
1922                         ktime_add(lock->l_last_used,
1923                                   ktime_set(ns->ns_dirty_age_limit, 0)))) {
1924                 unlock_res_and_lock(lock);
1925                 if (ldlm_bl_to_thread_lock(ns, NULL, lock))
1926                         ldlm_handle_bl_callback(ns, NULL, lock);
1927
1928                 EXIT;
1929                 return;
1930         }
1931         unlock_res_and_lock(lock);
1932         LDLM_LOCK_RELEASE(lock);
1933         EXIT;
1934 }
1935
1936 static int ldlm_callback_reply(struct ptlrpc_request *req, int rc)
1937 {
1938         if (req->rq_no_reply)
1939                 return 0;
1940
1941         req->rq_status = rc;
1942         if (!req->rq_packed_final) {
1943                 rc = lustre_pack_reply(req, 1, NULL, NULL);
1944                 if (rc)
1945                         return rc;
1946         }
1947         return ptlrpc_reply(req);
1948 }
1949
1950 static int __ldlm_bl_to_thread(struct ldlm_bl_work_item *blwi,
1951                                enum ldlm_cancel_flags cancel_flags)
1952 {
1953         struct ldlm_bl_pool *blp = ldlm_state->ldlm_bl_pool;
1954         ENTRY;
1955
1956         spin_lock(&blp->blp_lock);
1957         if (blwi->blwi_lock &&
1958             ldlm_is_discard_data(blwi->blwi_lock)) {
1959                 /* add LDLM_FL_DISCARD_DATA requests to the priority list */
1960                 list_add_tail(&blwi->blwi_entry, &blp->blp_prio_list);
1961         } else {
1962                 /* other blocking callbacks are added to the regular list */
1963                 list_add_tail(&blwi->blwi_entry, &blp->blp_list);
1964         }
1965         spin_unlock(&blp->blp_lock);
1966
1967         wake_up(&blp->blp_waitq);
1968
1969         /* can not check blwi->blwi_flags as blwi could be already freed in
1970            LCF_ASYNC mode */
1971         if (!(cancel_flags & LCF_ASYNC))
1972                 wait_for_completion(&blwi->blwi_comp);
1973
1974         RETURN(0);
1975 }
1976
1977 static inline void init_blwi(struct ldlm_bl_work_item *blwi,
1978                              struct ldlm_namespace *ns,
1979                              struct ldlm_lock_desc *ld,
1980                              struct list_head *cancels, int count,
1981                              struct ldlm_lock *lock,
1982                              enum ldlm_cancel_flags cancel_flags)
1983 {
1984         init_completion(&blwi->blwi_comp);
1985         INIT_LIST_HEAD(&blwi->blwi_head);
1986
1987         if (memory_pressure_get())
1988                 blwi->blwi_mem_pressure = 1;
1989
1990         blwi->blwi_ns = ns;
1991         blwi->blwi_flags = cancel_flags;
1992         if (ld != NULL)
1993                 blwi->blwi_ld = *ld;
1994         if (count) {
1995                 list_add(&blwi->blwi_head, cancels);
1996                 list_del_init(cancels);
1997                 blwi->blwi_count = count;
1998         } else {
1999                 blwi->blwi_lock = lock;
2000         }
2001 }
2002
2003 /**
2004  * Queues a list of locks \a cancels containing \a count locks
2005  * for later processing by a blocking thread.  If \a count is zero,
2006  * then the lock referenced as \a lock is queued instead.
2007  *
2008  * The blocking thread would then call ->l_blocking_ast callback in the lock.
2009  * If list addition fails an error is returned and caller is supposed to
2010  * call ->l_blocking_ast itself.
2011  */
2012 static int ldlm_bl_to_thread(struct ldlm_namespace *ns,
2013                              struct ldlm_lock_desc *ld,
2014                              struct ldlm_lock *lock,
2015                              struct list_head *cancels, int count,
2016                              enum ldlm_cancel_flags cancel_flags)
2017 {
2018         ENTRY;
2019
2020         if (cancels && count == 0)
2021                 RETURN(0);
2022
2023         if (cancel_flags & LCF_ASYNC) {
2024                 struct ldlm_bl_work_item *blwi;
2025
2026                 OBD_ALLOC(blwi, sizeof(*blwi));
2027                 if (blwi == NULL)
2028                         RETURN(-ENOMEM);
2029                 init_blwi(blwi, ns, ld, cancels, count, lock, cancel_flags);
2030
2031                 RETURN(__ldlm_bl_to_thread(blwi, cancel_flags));
2032         } else {
2033                 /* if it is synchronous call do minimum mem alloc, as it could
2034                  * be triggered from kernel shrinker
2035                  */
2036                 struct ldlm_bl_work_item blwi;
2037
2038                 memset(&blwi, 0, sizeof(blwi));
2039                 init_blwi(&blwi, ns, ld, cancels, count, lock, cancel_flags);
2040                 RETURN(__ldlm_bl_to_thread(&blwi, cancel_flags));
2041         }
2042 }
2043
2044
2045 int ldlm_bl_to_thread_lock(struct ldlm_namespace *ns, struct ldlm_lock_desc *ld,
2046                            struct ldlm_lock *lock)
2047 {
2048         return ldlm_bl_to_thread(ns, ld, lock, NULL, 0, LCF_ASYNC);
2049 }
2050
2051 int ldlm_bl_to_thread_list(struct ldlm_namespace *ns, struct ldlm_lock_desc *ld,
2052                            struct list_head *cancels, int count,
2053                            enum ldlm_cancel_flags cancel_flags)
2054 {
2055         return ldlm_bl_to_thread(ns, ld, NULL, cancels, count, cancel_flags);
2056 }
2057
2058 int ldlm_bl_thread_wakeup(void)
2059 {
2060         wake_up(&ldlm_state->ldlm_bl_pool->blp_waitq);
2061         return 0;
2062 }
2063
2064 /* Setinfo coming from Server (eg MDT) to Client (eg MDC)! */
2065 static int ldlm_handle_setinfo(struct ptlrpc_request *req)
2066 {
2067         struct obd_device *obd = req->rq_export->exp_obd;
2068         char *key;
2069         void *val;
2070         int keylen, vallen;
2071         int rc = -ENOSYS;
2072         ENTRY;
2073
2074         DEBUG_REQ(D_HSM, req, "%s: handle setinfo\n", obd->obd_name);
2075
2076         req_capsule_set(&req->rq_pill, &RQF_OBD_SET_INFO);
2077
2078         key = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_KEY);
2079         if (key == NULL) {
2080                 DEBUG_REQ(D_IOCTL, req, "no set_info key");
2081                 RETURN(-EFAULT);
2082         }
2083         keylen = req_capsule_get_size(&req->rq_pill, &RMF_SETINFO_KEY,
2084                                       RCL_CLIENT);
2085         val = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_VAL);
2086         if (val == NULL) {
2087                 DEBUG_REQ(D_IOCTL, req, "no set_info val");
2088                 RETURN(-EFAULT);
2089         }
2090         vallen = req_capsule_get_size(&req->rq_pill, &RMF_SETINFO_VAL,
2091                                       RCL_CLIENT);
2092
2093         /* We are responsible for swabbing contents of val */
2094
2095         if (KEY_IS(KEY_HSM_COPYTOOL_SEND))
2096                 /* Pass it on to mdc (the "export" in this case) */
2097                 rc = obd_set_info_async(req->rq_svc_thread->t_env,
2098                                         req->rq_export,
2099                                         sizeof(KEY_HSM_COPYTOOL_SEND),
2100                                         KEY_HSM_COPYTOOL_SEND,
2101                                         vallen, val, NULL);
2102         else
2103                 DEBUG_REQ(D_WARNING, req, "ignoring unknown key %s", key);
2104
2105         return rc;
2106 }
2107
2108 static inline void ldlm_callback_errmsg(struct ptlrpc_request *req,
2109                                         const char *msg, int rc,
2110                                         const struct lustre_handle *handle)
2111 {
2112         DEBUG_REQ((req->rq_no_reply || rc) ? D_WARNING : D_DLMTRACE, req,
2113                   "%s: [nid %s] [rc %d] [lock %#llx]",
2114                   msg, libcfs_id2str(req->rq_peer), rc,
2115                   handle ? handle->cookie : 0);
2116         if (req->rq_no_reply)
2117                 CWARN("No reply was sent, maybe cause bug 21636.\n");
2118         else if (rc)
2119                 CWARN("Send reply failed, maybe cause bug 21636.\n");
2120 }
2121
2122 /* TODO: handle requests in a similar way as MDT: see mdt_handle_common() */
2123 static int ldlm_callback_handler(struct ptlrpc_request *req)
2124 {
2125         struct ldlm_namespace *ns;
2126         struct ldlm_request *dlm_req;
2127         struct ldlm_lock *lock;
2128         int rc;
2129         ENTRY;
2130
2131         /* Requests arrive in sender's byte order.  The ptlrpc service
2132          * handler has already checked and, if necessary, byte-swapped the
2133          * incoming request message body, but I am responsible for the
2134          * message buffers. */
2135
2136         /* do nothing for sec context finalize */
2137         if (lustre_msg_get_opc(req->rq_reqmsg) == SEC_CTX_FINI)
2138                 RETURN(0);
2139
2140         req_capsule_init(&req->rq_pill, req, RCL_SERVER);
2141
2142         if (req->rq_export == NULL) {
2143                 rc = ldlm_callback_reply(req, -ENOTCONN);
2144                 ldlm_callback_errmsg(req, "Operate on unconnected server",
2145                                      rc, NULL);
2146                 RETURN(0);
2147         }
2148
2149         LASSERT(req->rq_export != NULL);
2150         LASSERT(req->rq_export->exp_obd != NULL);
2151
2152         switch (lustre_msg_get_opc(req->rq_reqmsg)) {
2153         case LDLM_BL_CALLBACK:
2154                 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_BL_CALLBACK_NET)) {
2155                         if (cfs_fail_err)
2156                                 ldlm_callback_reply(req, -(int)cfs_fail_err);
2157                         RETURN(0);
2158                 }
2159                 break;
2160         case LDLM_CP_CALLBACK:
2161                 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CP_CALLBACK_NET))
2162                         RETURN(0);
2163                 break;
2164         case LDLM_GL_CALLBACK:
2165                 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_GL_CALLBACK_NET))
2166                         RETURN(0);
2167                 break;
2168         case LDLM_SET_INFO:
2169                 rc = ldlm_handle_setinfo(req);
2170                 ldlm_callback_reply(req, rc);
2171                 RETURN(0);
2172         default:
2173                 CERROR("unknown opcode %u\n",
2174                        lustre_msg_get_opc(req->rq_reqmsg));
2175                 ldlm_callback_reply(req, -EPROTO);
2176                 RETURN(0);
2177         }
2178
2179         ns = req->rq_export->exp_obd->obd_namespace;
2180         LASSERT(ns != NULL);
2181
2182         req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK);
2183
2184         dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2185         if (dlm_req == NULL) {
2186                 rc = ldlm_callback_reply(req, -EPROTO);
2187                 ldlm_callback_errmsg(req, "Operate without parameter", rc,
2188                                      NULL);
2189                 RETURN(0);
2190         }
2191
2192         /* Force a known safe race, send a cancel to the server for a lock
2193          * which the server has already started a blocking callback on. */
2194         if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_BL_CB_RACE) &&
2195             lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK) {
2196                 rc = ldlm_cli_cancel(&dlm_req->lock_handle[0], 0);
2197                 if (rc < 0)
2198                         CERROR("ldlm_cli_cancel: %d\n", rc);
2199         }
2200
2201         lock = ldlm_handle2lock_long(&dlm_req->lock_handle[0], 0);
2202         if (!lock) {
2203                 CDEBUG(D_DLMTRACE, "callback on lock %#llx - lock "
2204                        "disappeared\n", dlm_req->lock_handle[0].cookie);
2205                 rc = ldlm_callback_reply(req, -EINVAL);
2206                 ldlm_callback_errmsg(req, "Operate with invalid parameter", rc,
2207                                      &dlm_req->lock_handle[0]);
2208                 RETURN(0);
2209         }
2210
2211         if (ldlm_is_fail_loc(lock) &&
2212             lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK)
2213                 OBD_RACE(OBD_FAIL_LDLM_CP_BL_RACE);
2214
2215         /* Copy hints/flags (e.g. LDLM_FL_DISCARD_DATA) from AST. */
2216         lock_res_and_lock(lock);
2217         lock->l_flags |= ldlm_flags_from_wire(dlm_req->lock_flags &
2218                                               LDLM_FL_AST_MASK);
2219         if (lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK) {
2220                 /* If somebody cancels lock and cache is already dropped,
2221                  * or lock is failed before cp_ast received on client,
2222                  * we can tell the server we have no lock. Otherwise, we
2223                  * should send cancel after dropping the cache. */
2224                 if ((ldlm_is_canceling(lock) && ldlm_is_bl_done(lock)) ||
2225                      ldlm_is_failed(lock)) {
2226                         LDLM_DEBUG(lock, "callback on lock %llx - lock disappeared",
2227                                    dlm_req->lock_handle[0].cookie);
2228                         unlock_res_and_lock(lock);
2229                         LDLM_LOCK_RELEASE(lock);
2230                         rc = ldlm_callback_reply(req, -EINVAL);
2231                         ldlm_callback_errmsg(req, "Operate on stale lock", rc,
2232                                              &dlm_req->lock_handle[0]);
2233                         RETURN(0);
2234                 }
2235                 /* BL_AST locks are not needed in LRU.
2236                  * Let ldlm_cancel_lru() be fast. */
2237                 ldlm_lock_remove_from_lru(lock);
2238                 ldlm_set_bl_ast(lock);
2239         }
2240         unlock_res_and_lock(lock);
2241
2242         /* We want the ost thread to get this reply so that it can respond
2243          * to ost requests (write cache writeback) that might be triggered
2244          * in the callback.
2245          *
2246          * But we'd also like to be able to indicate in the reply that we're
2247          * cancelling right now, because it's unused, or have an intent result
2248          * in the reply, so we might have to push the responsibility for sending
2249          * the reply down into the AST handlers, alas. */
2250
2251         switch (lustre_msg_get_opc(req->rq_reqmsg)) {
2252         case LDLM_BL_CALLBACK:
2253                 CDEBUG(D_INODE, "blocking ast\n");
2254                 req_capsule_extend(&req->rq_pill, &RQF_LDLM_BL_CALLBACK);
2255                 if (!ldlm_is_cancel_on_block(lock)) {
2256                         rc = ldlm_callback_reply(req, 0);
2257                         if (req->rq_no_reply || rc)
2258                                 ldlm_callback_errmsg(req, "Normal process", rc,
2259                                                      &dlm_req->lock_handle[0]);
2260                 }
2261                 if (ldlm_bl_to_thread_lock(ns, &dlm_req->lock_desc, lock))
2262                         ldlm_handle_bl_callback(ns, &dlm_req->lock_desc, lock);
2263                 break;
2264         case LDLM_CP_CALLBACK:
2265                 CDEBUG(D_INODE, "completion ast\n");
2266                 req_capsule_extend(&req->rq_pill, &RQF_LDLM_CP_CALLBACK);
2267                 ldlm_callback_reply(req, 0);
2268                 ldlm_handle_cp_callback(req, ns, dlm_req, lock);
2269                 break;
2270         case LDLM_GL_CALLBACK:
2271                 CDEBUG(D_INODE, "glimpse ast\n");
2272                 req_capsule_extend(&req->rq_pill, &RQF_LDLM_GL_CALLBACK);
2273                 ldlm_handle_gl_callback(req, ns, dlm_req, lock);
2274                 break;
2275         default:
2276                 LBUG();                         /* checked above */
2277         }
2278
2279         RETURN(0);
2280 }
2281
2282 #ifdef HAVE_SERVER_SUPPORT
2283 /**
2284  * Main handler for canceld thread.
2285  *
2286  * Separated into its own thread to avoid deadlocks.
2287  */
2288 static int ldlm_cancel_handler(struct ptlrpc_request *req)
2289 {
2290         int rc;
2291
2292         ENTRY;
2293
2294         /* Requests arrive in sender's byte order.  The ptlrpc service
2295          * handler has already checked and, if necessary, byte-swapped the
2296          * incoming request message body, but I am responsible for the
2297          * message buffers. */
2298
2299         req_capsule_init(&req->rq_pill, req, RCL_SERVER);
2300
2301         if (req->rq_export == NULL) {
2302                 struct ldlm_request *dlm_req;
2303
2304                 CERROR("%s from %s arrived at %lu with bad export cookie "
2305                        "%llu\n",
2306                        ll_opcode2str(lustre_msg_get_opc(req->rq_reqmsg)),
2307                        libcfs_nid2str(req->rq_peer.nid),
2308                        req->rq_arrival_time.tv_sec,
2309                        lustre_msg_get_handle(req->rq_reqmsg)->cookie);
2310
2311                 if (lustre_msg_get_opc(req->rq_reqmsg) == LDLM_CANCEL) {
2312                         req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK);
2313                         dlm_req = req_capsule_client_get(&req->rq_pill,
2314                                                          &RMF_DLM_REQ);
2315                         if (dlm_req != NULL)
2316                                 ldlm_lock_dump_handle(D_ERROR,
2317                                                       &dlm_req->lock_handle[0]);
2318                 }
2319                 ldlm_callback_reply(req, -ENOTCONN);
2320                 RETURN(0);
2321         }
2322
2323         switch (lustre_msg_get_opc(req->rq_reqmsg)) {
2324         /* XXX FIXME move this back to mds/handler.c, bug 249 */
2325         case LDLM_CANCEL:
2326                 req_capsule_set(&req->rq_pill, &RQF_LDLM_CANCEL);
2327                 CDEBUG(D_INODE, "cancel\n");
2328                 if (CFS_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_NET) ||
2329                     CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_CANCEL_RESEND) ||
2330                     CFS_FAIL_CHECK(OBD_FAIL_LDLM_BL_EVICT))
2331                         RETURN(0);
2332                 rc = ldlm_handle_cancel(req);
2333                 break;
2334         case LDLM_CONVERT:
2335         {
2336                 struct ldlm_request *dlm_req;
2337
2338                 req_capsule_set(&req->rq_pill, &RQF_LDLM_CONVERT);
2339                 CDEBUG(D_INODE, "convert\n");
2340
2341                 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2342                 if (dlm_req == NULL) {
2343                         CDEBUG(D_INFO, "bad request buffer for cancel\n");
2344                         rc = ldlm_callback_reply(req, -EPROTO);
2345                 } else {
2346                         req->rq_status = ldlm_handle_convert0(req, dlm_req);
2347                         rc = ptlrpc_reply(req);
2348                 }
2349                 break;
2350         }
2351         default:
2352                 CERROR("invalid opcode %d\n",
2353                        lustre_msg_get_opc(req->rq_reqmsg));
2354                 req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK);
2355                 rc = ldlm_callback_reply(req, -EINVAL);
2356         }
2357
2358         RETURN(rc);
2359 }
2360
2361 static int ldlm_cancel_hpreq_lock_match(struct ptlrpc_request *req,
2362                                         struct ldlm_lock *lock)
2363 {
2364         struct ldlm_request *dlm_req;
2365         struct lustre_handle lockh;
2366         int rc = 0;
2367         int i;
2368
2369         ENTRY;
2370
2371         dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2372         if (dlm_req == NULL)
2373                 RETURN(0);
2374
2375         ldlm_lock2handle(lock, &lockh);
2376         for (i = 0; i < dlm_req->lock_count; i++) {
2377                 if (lustre_handle_equal(&dlm_req->lock_handle[i],
2378                                         &lockh)) {
2379                         DEBUG_REQ(D_RPCTRACE, req,
2380                                   "Prio raised by lock %#llx.", lockh.cookie);
2381                         rc = 1;
2382                         break;
2383                 }
2384         }
2385
2386         RETURN(rc);
2387 }
2388
2389 static int ldlm_cancel_hpreq_check(struct ptlrpc_request *req)
2390 {
2391         struct ldlm_request *dlm_req;
2392         int rc = 0;
2393         int i;
2394
2395         ENTRY;
2396
2397         /* no prolong in recovery */
2398         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
2399                 RETURN(0);
2400
2401         dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2402         if (dlm_req == NULL)
2403                 RETURN(-EFAULT);
2404
2405         for (i = 0; i < dlm_req->lock_count; i++) {
2406                 struct ldlm_lock *lock;
2407
2408                 lock = ldlm_handle2lock(&dlm_req->lock_handle[i]);
2409                 if (lock == NULL)
2410                         continue;
2411
2412                 rc = ldlm_is_ast_sent(lock) ? 1 : 0;
2413                 if (rc)
2414                         LDLM_DEBUG(lock, "hpreq cancel/convert lock");
2415                 LDLM_LOCK_PUT(lock);
2416
2417                 if (rc)
2418                         break;
2419         }
2420
2421         RETURN(rc);
2422 }
2423
2424 static struct ptlrpc_hpreq_ops ldlm_cancel_hpreq_ops = {
2425         .hpreq_lock_match = ldlm_cancel_hpreq_lock_match,
2426         .hpreq_check      = ldlm_cancel_hpreq_check,
2427         .hpreq_fini       = NULL,
2428 };
2429
2430 static int ldlm_hpreq_handler(struct ptlrpc_request *req)
2431 {
2432         ENTRY;
2433
2434         req_capsule_init(&req->rq_pill, req, RCL_SERVER);
2435
2436         if (req->rq_export == NULL)
2437                 RETURN(0);
2438
2439         if (LDLM_CANCEL == lustre_msg_get_opc(req->rq_reqmsg)) {
2440                 req_capsule_set(&req->rq_pill, &RQF_LDLM_CANCEL);
2441                 req->rq_ops = &ldlm_cancel_hpreq_ops;
2442         } else if (LDLM_CONVERT == lustre_msg_get_opc(req->rq_reqmsg)) {
2443                 req_capsule_set(&req->rq_pill, &RQF_LDLM_CONVERT);
2444                 req->rq_ops = &ldlm_cancel_hpreq_ops;
2445         }
2446         RETURN(0);
2447 }
2448
2449 static int ldlm_revoke_lock_cb(struct cfs_hash *hs, struct cfs_hash_bd *bd,
2450                                struct hlist_node *hnode, void *data)
2451
2452 {
2453         struct list_head         *rpc_list = data;
2454         struct ldlm_lock   *lock = cfs_hash_object(hs, hnode);
2455
2456         lock_res_and_lock(lock);
2457
2458         if (lock->l_req_mode != lock->l_granted_mode) {
2459                 unlock_res_and_lock(lock);
2460                 return 0;
2461         }
2462
2463         LASSERT(lock->l_resource);
2464         if (lock->l_resource->lr_type != LDLM_IBITS &&
2465             lock->l_resource->lr_type != LDLM_PLAIN) {
2466                 unlock_res_and_lock(lock);
2467                 return 0;
2468         }
2469
2470         if (ldlm_is_ast_sent(lock)) {
2471                 unlock_res_and_lock(lock);
2472                 return 0;
2473         }
2474
2475         LASSERT(lock->l_blocking_ast);
2476         LASSERT(!lock->l_blocking_lock);
2477
2478         ldlm_set_ast_sent(lock);
2479         if (lock->l_export && lock->l_export->exp_lock_hash) {
2480                 /* NB: it's safe to call cfs_hash_del() even lock isn't
2481                  * in exp_lock_hash. */
2482                 /* In the function below, .hs_keycmp resolves to
2483                  * ldlm_export_lock_keycmp() */
2484                 /* coverity[overrun-buffer-val] */
2485                 cfs_hash_del(lock->l_export->exp_lock_hash,
2486                              &lock->l_remote_handle, &lock->l_exp_hash);
2487         }
2488
2489         list_add_tail(&lock->l_rk_ast, rpc_list);
2490         LDLM_LOCK_GET(lock);
2491
2492         unlock_res_and_lock(lock);
2493         return 0;
2494 }
2495
2496 void ldlm_revoke_export_locks(struct obd_export *exp)
2497 {
2498         struct list_head  rpc_list;
2499         ENTRY;
2500
2501         INIT_LIST_HEAD(&rpc_list);
2502         cfs_hash_for_each_nolock(exp->exp_lock_hash,
2503                                  ldlm_revoke_lock_cb, &rpc_list, 0);
2504         ldlm_run_ast_work(exp->exp_obd->obd_namespace, &rpc_list,
2505                           LDLM_WORK_REVOKE_AST);
2506
2507         EXIT;
2508 }
2509 EXPORT_SYMBOL(ldlm_revoke_export_locks);
2510 #endif /* HAVE_SERVER_SUPPORT */
2511
2512 static int ldlm_bl_get_work(struct ldlm_bl_pool *blp,
2513                             struct ldlm_bl_work_item **p_blwi,
2514                             struct obd_export **p_exp)
2515 {
2516         struct ldlm_bl_work_item *blwi = NULL;
2517         static unsigned int num_bl = 0;
2518         static unsigned int num_stale;
2519         int num_th = atomic_read(&blp->blp_num_threads);
2520
2521         *p_exp = obd_stale_export_get();
2522
2523         spin_lock(&blp->blp_lock);
2524         if (*p_exp != NULL) {
2525                 if (num_th == 1 || ++num_stale < num_th) {
2526                         spin_unlock(&blp->blp_lock);
2527                         return 1;
2528                 } else {
2529                         num_stale = 0;
2530                 }
2531         }
2532
2533         /* process a request from the blp_list at least every blp_num_threads */
2534         if (!list_empty(&blp->blp_list) &&
2535             (list_empty(&blp->blp_prio_list) || num_bl == 0))
2536                 blwi = list_entry(blp->blp_list.next,
2537                                   struct ldlm_bl_work_item, blwi_entry);
2538         else
2539                 if (!list_empty(&blp->blp_prio_list))
2540                         blwi = list_entry(blp->blp_prio_list.next,
2541                                           struct ldlm_bl_work_item,
2542                                           blwi_entry);
2543
2544         if (blwi) {
2545                 if (++num_bl >= num_th)
2546                         num_bl = 0;
2547                 list_del(&blwi->blwi_entry);
2548         }
2549         spin_unlock(&blp->blp_lock);
2550         *p_blwi = blwi;
2551
2552         if (*p_exp != NULL && *p_blwi != NULL) {
2553                 obd_stale_export_put(*p_exp);
2554                 *p_exp = NULL;
2555         }
2556
2557         return (*p_blwi != NULL || *p_exp != NULL) ? 1 : 0;
2558 }
2559
2560 /* This only contains temporary data until the thread starts */
2561 struct ldlm_bl_thread_data {
2562         struct ldlm_bl_pool     *bltd_blp;
2563         struct completion       bltd_comp;
2564         int                     bltd_num;
2565 };
2566
2567 static int ldlm_bl_thread_main(void *arg);
2568
2569 static int ldlm_bl_thread_start(struct ldlm_bl_pool *blp, bool check_busy)
2570 {
2571         struct ldlm_bl_thread_data bltd = { .bltd_blp = blp };
2572         struct task_struct *task;
2573
2574         init_completion(&bltd.bltd_comp);
2575
2576         bltd.bltd_num = atomic_inc_return(&blp->blp_num_threads);
2577         if (bltd.bltd_num >= blp->blp_max_threads) {
2578                 atomic_dec(&blp->blp_num_threads);
2579                 return 0;
2580         }
2581
2582         LASSERTF(bltd.bltd_num > 0, "thread num:%d\n", bltd.bltd_num);
2583         if (check_busy &&
2584             atomic_read(&blp->blp_busy_threads) < (bltd.bltd_num - 1)) {
2585                 atomic_dec(&blp->blp_num_threads);
2586                 return 0;
2587         }
2588
2589         task = kthread_run(ldlm_bl_thread_main, &bltd, "ldlm_bl_%02d",
2590                            bltd.bltd_num);
2591         if (IS_ERR(task)) {
2592                 CERROR("cannot start LDLM thread ldlm_bl_%02d: rc %ld\n",
2593                        bltd.bltd_num, PTR_ERR(task));
2594                 atomic_dec(&blp->blp_num_threads);
2595                 return PTR_ERR(task);
2596         }
2597         wait_for_completion(&bltd.bltd_comp);
2598
2599         return 0;
2600 }
2601
2602 /* Not fatal if racy and have a few too many threads */
2603 static int ldlm_bl_thread_need_create(struct ldlm_bl_pool *blp,
2604                                       struct ldlm_bl_work_item *blwi)
2605 {
2606         if (atomic_read(&blp->blp_num_threads) >= blp->blp_max_threads)
2607                 return 0;
2608
2609         if (atomic_read(&blp->blp_busy_threads) <
2610             atomic_read(&blp->blp_num_threads))
2611                 return 0;
2612
2613         if (blwi != NULL && (blwi->blwi_ns == NULL ||
2614                              blwi->blwi_mem_pressure))
2615                 return 0;
2616
2617         return 1;
2618 }
2619
2620 static int ldlm_bl_thread_blwi(struct ldlm_bl_pool *blp,
2621                                struct ldlm_bl_work_item *blwi)
2622 {
2623         ENTRY;
2624
2625         if (blwi->blwi_ns == NULL)
2626                 /* added by ldlm_cleanup() */
2627                 RETURN(LDLM_ITER_STOP);
2628
2629         if (blwi->blwi_mem_pressure)
2630                 memory_pressure_set();
2631
2632         OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_PAUSE_CANCEL2, 4);
2633
2634         if (blwi->blwi_count) {
2635                 int count;
2636                 /* The special case when we cancel locks in lru
2637                  * asynchronously, we pass the list of locks here.
2638                  * Thus locks are marked LDLM_FL_CANCELING, but NOT
2639                  * canceled locally yet. */
2640                 count = ldlm_cli_cancel_list_local(&blwi->blwi_head,
2641                                                    blwi->blwi_count,
2642                                                    LCF_BL_AST);
2643                 ldlm_cli_cancel_list(&blwi->blwi_head, count, NULL,
2644                                      blwi->blwi_flags);
2645         } else {
2646                 ldlm_handle_bl_callback(blwi->blwi_ns, &blwi->blwi_ld,
2647                                         blwi->blwi_lock);
2648         }
2649         if (blwi->blwi_mem_pressure)
2650                 memory_pressure_clr();
2651
2652         if (blwi->blwi_flags & LCF_ASYNC)
2653                 OBD_FREE(blwi, sizeof(*blwi));
2654         else
2655                 complete(&blwi->blwi_comp);
2656
2657         RETURN(0);
2658 }
2659
2660 /**
2661  * Cancel stale locks on export. Cancel blocked locks first.
2662  * If the given export has blocked locks, the next in the list may have
2663  * them too, thus cancel not blocked locks only if the current export has
2664  * no blocked locks.
2665  **/
2666 static int ldlm_bl_thread_exports(struct ldlm_bl_pool *blp,
2667                                   struct obd_export *exp)
2668 {
2669         int num;
2670         ENTRY;
2671
2672         OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_BL_EVICT, 4);
2673
2674         num = ldlm_export_cancel_blocked_locks(exp);
2675         if (num == 0)
2676                 ldlm_export_cancel_locks(exp);
2677
2678         obd_stale_export_put(exp);
2679
2680         RETURN(0);
2681 }
2682
2683
2684 /**
2685  * Main blocking requests processing thread.
2686  *
2687  * Callers put locks into its queue by calling ldlm_bl_to_thread.
2688  * This thread in the end ends up doing actual call to ->l_blocking_ast
2689  * for queued locks.
2690  */
2691 static int ldlm_bl_thread_main(void *arg)
2692 {
2693         struct ldlm_bl_pool *blp;
2694         struct ldlm_bl_thread_data *bltd = arg;
2695         ENTRY;
2696
2697         blp = bltd->bltd_blp;
2698
2699         complete(&bltd->bltd_comp);
2700         /* cannot use bltd after this, it is only on caller's stack */
2701
2702         while (1) {
2703                 struct l_wait_info lwi = { 0 };
2704                 struct ldlm_bl_work_item *blwi = NULL;
2705                 struct obd_export *exp = NULL;
2706                 int rc;
2707
2708                 rc = ldlm_bl_get_work(blp, &blwi, &exp);
2709
2710                 if (rc == 0)
2711                         l_wait_event_exclusive(blp->blp_waitq,
2712                                                ldlm_bl_get_work(blp, &blwi,
2713                                                                 &exp),
2714                                                &lwi);
2715                 atomic_inc(&blp->blp_busy_threads);
2716
2717                 if (ldlm_bl_thread_need_create(blp, blwi))
2718                         /* discard the return value, we tried */
2719                         ldlm_bl_thread_start(blp, true);
2720
2721                 if (exp)
2722                         rc = ldlm_bl_thread_exports(blp, exp);
2723                 else if (blwi)
2724                         rc = ldlm_bl_thread_blwi(blp, blwi);
2725
2726                 atomic_dec(&blp->blp_busy_threads);
2727
2728                 if (rc == LDLM_ITER_STOP)
2729                         break;
2730
2731                 /* If there are many namespaces, we will not sleep waiting for
2732                  * work, and must do a cond_resched to avoid holding the CPU
2733                  * for too long */
2734                 cond_resched();
2735         }
2736
2737         atomic_dec(&blp->blp_num_threads);
2738         complete(&blp->blp_comp);
2739         RETURN(0);
2740 }
2741
2742
2743 static int ldlm_setup(void);
2744 static int ldlm_cleanup(void);
2745
2746 int ldlm_get_ref(void)
2747 {
2748         int rc = 0;
2749         ENTRY;
2750         mutex_lock(&ldlm_ref_mutex);
2751         if (++ldlm_refcount == 1) {
2752                 rc = ldlm_setup();
2753                 if (rc)
2754                         ldlm_refcount--;
2755         }
2756         mutex_unlock(&ldlm_ref_mutex);
2757
2758         RETURN(rc);
2759 }
2760
2761 void ldlm_put_ref(void)
2762 {
2763         ENTRY;
2764         mutex_lock(&ldlm_ref_mutex);
2765         if (ldlm_refcount == 1) {
2766                 int rc = ldlm_cleanup();
2767                 if (rc)
2768                         CERROR("ldlm_cleanup failed: %d\n", rc);
2769                 else
2770                         ldlm_refcount--;
2771         } else {
2772                 ldlm_refcount--;
2773         }
2774         mutex_unlock(&ldlm_ref_mutex);
2775
2776         EXIT;
2777 }
2778
2779 /*
2780  * Export handle<->lock hash operations.
2781  */
2782 static unsigned
2783 ldlm_export_lock_hash(struct cfs_hash *hs, const void *key, unsigned mask)
2784 {
2785         return cfs_hash_u64_hash(((struct lustre_handle *)key)->cookie, mask);
2786 }
2787
2788 static void *
2789 ldlm_export_lock_key(struct hlist_node *hnode)
2790 {
2791         struct ldlm_lock *lock;
2792
2793         lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2794         return &lock->l_remote_handle;
2795 }
2796
2797 static void
2798 ldlm_export_lock_keycpy(struct hlist_node *hnode, void *key)
2799 {
2800         struct ldlm_lock     *lock;
2801
2802         lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2803         lock->l_remote_handle = *(struct lustre_handle *)key;
2804 }
2805
2806 static int
2807 ldlm_export_lock_keycmp(const void *key, struct hlist_node *hnode)
2808 {
2809         return lustre_handle_equal(ldlm_export_lock_key(hnode), key);
2810 }
2811
2812 static void *
2813 ldlm_export_lock_object(struct hlist_node *hnode)
2814 {
2815         return hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2816 }
2817
2818 static void
2819 ldlm_export_lock_get(struct cfs_hash *hs, struct hlist_node *hnode)
2820 {
2821         struct ldlm_lock *lock;
2822
2823         lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2824         LDLM_LOCK_GET(lock);
2825 }
2826
2827 static void
2828 ldlm_export_lock_put(struct cfs_hash *hs, struct hlist_node *hnode)
2829 {
2830         struct ldlm_lock *lock;
2831
2832         lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2833         LDLM_LOCK_RELEASE(lock);
2834 }
2835
2836 static struct cfs_hash_ops ldlm_export_lock_ops = {
2837         .hs_hash        = ldlm_export_lock_hash,
2838         .hs_key         = ldlm_export_lock_key,
2839         .hs_keycmp      = ldlm_export_lock_keycmp,
2840         .hs_keycpy      = ldlm_export_lock_keycpy,
2841         .hs_object      = ldlm_export_lock_object,
2842         .hs_get         = ldlm_export_lock_get,
2843         .hs_put         = ldlm_export_lock_put,
2844         .hs_put_locked  = ldlm_export_lock_put,
2845 };
2846
2847 int ldlm_init_export(struct obd_export *exp)
2848 {
2849         int rc;
2850         ENTRY;
2851
2852         exp->exp_lock_hash =
2853                 cfs_hash_create(obd_uuid2str(&exp->exp_client_uuid),
2854                                 HASH_EXP_LOCK_CUR_BITS,
2855                                 HASH_EXP_LOCK_MAX_BITS,
2856                                 HASH_EXP_LOCK_BKT_BITS, 0,
2857                                 CFS_HASH_MIN_THETA, CFS_HASH_MAX_THETA,
2858                                 &ldlm_export_lock_ops,
2859                                 CFS_HASH_DEFAULT | CFS_HASH_REHASH_KEY |
2860                                 CFS_HASH_NBLK_CHANGE);
2861
2862         if (!exp->exp_lock_hash)
2863                 RETURN(-ENOMEM);
2864
2865         rc = ldlm_init_flock_export(exp);
2866         if (rc)
2867                 GOTO(err, rc);
2868
2869         RETURN(0);
2870 err:
2871         ldlm_destroy_export(exp);
2872         RETURN(rc);
2873 }
2874 EXPORT_SYMBOL(ldlm_init_export);
2875
2876 void ldlm_destroy_export(struct obd_export *exp)
2877 {
2878         ENTRY;
2879         cfs_hash_putref(exp->exp_lock_hash);
2880         exp->exp_lock_hash = NULL;
2881
2882         ldlm_destroy_flock_export(exp);
2883         EXIT;
2884 }
2885 EXPORT_SYMBOL(ldlm_destroy_export);
2886
2887 static ssize_t cancel_unused_locks_before_replay_show(struct kobject *kobj,
2888                                                       struct attribute *attr,
2889                                                       char *buf)
2890 {
2891         return sprintf(buf, "%d\n", ldlm_cancel_unused_locks_before_replay);
2892 }
2893
2894 static ssize_t cancel_unused_locks_before_replay_store(struct kobject *kobj,
2895                                                        struct attribute *attr,
2896                                                        const char *buffer,
2897                                                        size_t count)
2898 {
2899         int rc;
2900         unsigned long val;
2901
2902         rc = kstrtoul(buffer, 10, &val);
2903         if (rc)
2904                 return rc;
2905
2906         ldlm_cancel_unused_locks_before_replay = val;
2907
2908         return count;
2909 }
2910 LUSTRE_RW_ATTR(cancel_unused_locks_before_replay);
2911
2912 static struct attribute *ldlm_attrs[] = {
2913         &lustre_attr_cancel_unused_locks_before_replay.attr,
2914         NULL,
2915 };
2916
2917 static struct attribute_group ldlm_attr_group = {
2918         .attrs = ldlm_attrs,
2919 };
2920
2921 static int ldlm_setup(void)
2922 {
2923         static struct ptlrpc_service_conf       conf;
2924         struct ldlm_bl_pool                    *blp = NULL;
2925 #ifdef HAVE_SERVER_SUPPORT
2926         struct task_struct *task;
2927 #endif /* HAVE_SERVER_SUPPORT */
2928         int i;
2929         int rc = 0;
2930
2931         ENTRY;
2932
2933         if (ldlm_state != NULL)
2934                 RETURN(-EALREADY);
2935
2936         OBD_ALLOC(ldlm_state, sizeof(*ldlm_state));
2937         if (ldlm_state == NULL)
2938                 RETURN(-ENOMEM);
2939
2940         ldlm_kobj = kobject_create_and_add("ldlm", &lustre_kset->kobj);
2941         if (!ldlm_kobj)
2942                 GOTO(out, -ENOMEM);
2943
2944         rc = sysfs_create_group(ldlm_kobj, &ldlm_attr_group);
2945         if (rc)
2946                 GOTO(out, rc);
2947
2948         ldlm_ns_kset = kset_create_and_add("namespaces", NULL, ldlm_kobj);
2949         if (!ldlm_ns_kset)
2950                 GOTO(out, -ENOMEM);
2951
2952         ldlm_svc_kset = kset_create_and_add("services", NULL, ldlm_kobj);
2953         if (!ldlm_svc_kset)
2954                 GOTO(out, -ENOMEM);
2955
2956         rc = ldlm_debugfs_setup();
2957         if (rc != 0)
2958                 GOTO(out, rc);
2959
2960         memset(&conf, 0, sizeof(conf));
2961         conf = (typeof(conf)) {
2962                 .psc_name               = "ldlm_cbd",
2963                 .psc_watchdog_factor    = 2,
2964                 .psc_buf                = {
2965                         .bc_nbufs               = LDLM_CLIENT_NBUFS,
2966                         .bc_buf_size            = LDLM_BUFSIZE,
2967                         .bc_req_max_size        = LDLM_MAXREQSIZE,
2968                         .bc_rep_max_size        = LDLM_MAXREPSIZE,
2969                         .bc_req_portal          = LDLM_CB_REQUEST_PORTAL,
2970                         .bc_rep_portal          = LDLM_CB_REPLY_PORTAL,
2971                 },
2972                 .psc_thr                = {
2973                         .tc_thr_name            = "ldlm_cb",
2974                         .tc_thr_factor          = LDLM_THR_FACTOR,
2975                         .tc_nthrs_init          = LDLM_NTHRS_INIT,
2976                         .tc_nthrs_base          = LDLM_NTHRS_BASE,
2977                         .tc_nthrs_max           = LDLM_NTHRS_MAX,
2978                         .tc_nthrs_user          = ldlm_num_threads,
2979                         .tc_cpu_bind            = ldlm_cpu_bind,
2980                         .tc_ctx_tags            = LCT_MD_THREAD | LCT_DT_THREAD,
2981                 },
2982                 .psc_cpt                = {
2983                         .cc_pattern             = ldlm_cpts,
2984                         .cc_affinity            = true,
2985                 },
2986                 .psc_ops                = {
2987                         .so_req_handler         = ldlm_callback_handler,
2988                 },
2989         };
2990         ldlm_state->ldlm_cb_service = \
2991                         ptlrpc_register_service(&conf, ldlm_svc_kset,
2992                                                 ldlm_svc_debugfs_dir);
2993         if (IS_ERR(ldlm_state->ldlm_cb_service)) {
2994                 CERROR("failed to start service\n");
2995                 rc = PTR_ERR(ldlm_state->ldlm_cb_service);
2996                 ldlm_state->ldlm_cb_service = NULL;
2997                 GOTO(out, rc);
2998         }
2999
3000 #ifdef HAVE_SERVER_SUPPORT
3001         memset(&conf, 0, sizeof(conf));
3002         conf = (typeof(conf)) {
3003                 .psc_name               = "ldlm_canceld",
3004                 .psc_watchdog_factor    = 6,
3005                 .psc_buf                = {
3006                         .bc_nbufs               = LDLM_SERVER_NBUFS,
3007                         .bc_buf_size            = LDLM_BUFSIZE,
3008                         .bc_req_max_size        = LDLM_MAXREQSIZE,
3009                         .bc_rep_max_size        = LDLM_MAXREPSIZE,
3010                         .bc_req_portal          = LDLM_CANCEL_REQUEST_PORTAL,
3011                         .bc_rep_portal          = LDLM_CANCEL_REPLY_PORTAL,
3012
3013                 },
3014                 .psc_thr                = {
3015                         .tc_thr_name            = "ldlm_cn",
3016                         .tc_thr_factor          = LDLM_THR_FACTOR,
3017                         .tc_nthrs_init          = LDLM_NTHRS_INIT,
3018                         .tc_nthrs_base          = LDLM_NTHRS_BASE,
3019                         .tc_nthrs_max           = LDLM_NTHRS_MAX,
3020                         .tc_nthrs_user          = ldlm_num_threads,
3021                         .tc_cpu_bind            = ldlm_cpu_bind,
3022                         .tc_ctx_tags            = LCT_MD_THREAD | \
3023                                                   LCT_DT_THREAD | \
3024                                                   LCT_CL_THREAD,
3025                 },
3026                 .psc_cpt                = {
3027                         .cc_pattern             = ldlm_cpts,
3028                         .cc_affinity            = true,
3029                 },
3030                 .psc_ops                = {
3031                         .so_req_handler         = ldlm_cancel_handler,
3032                         .so_hpreq_handler       = ldlm_hpreq_handler,
3033                 },
3034         };
3035         ldlm_state->ldlm_cancel_service = \
3036                         ptlrpc_register_service(&conf, ldlm_svc_kset,
3037                                                 ldlm_svc_debugfs_dir);
3038         if (IS_ERR(ldlm_state->ldlm_cancel_service)) {
3039                 CERROR("failed to start service\n");
3040                 rc = PTR_ERR(ldlm_state->ldlm_cancel_service);
3041                 ldlm_state->ldlm_cancel_service = NULL;
3042                 GOTO(out, rc);
3043         }
3044 #endif /* HAVE_SERVER_SUPPORT */
3045
3046         OBD_ALLOC(blp, sizeof(*blp));
3047         if (blp == NULL)
3048                 GOTO(out, rc = -ENOMEM);
3049         ldlm_state->ldlm_bl_pool = blp;
3050
3051         spin_lock_init(&blp->blp_lock);
3052         INIT_LIST_HEAD(&blp->blp_list);
3053         INIT_LIST_HEAD(&blp->blp_prio_list);
3054         init_waitqueue_head(&blp->blp_waitq);
3055         atomic_set(&blp->blp_num_threads, 0);
3056         atomic_set(&blp->blp_busy_threads, 0);
3057
3058         if (ldlm_num_threads == 0) {
3059                 blp->blp_min_threads = LDLM_NTHRS_INIT;
3060                 blp->blp_max_threads = LDLM_NTHRS_MAX;
3061         } else {
3062                 blp->blp_min_threads = blp->blp_max_threads = \
3063                         min_t(int, LDLM_NTHRS_MAX, max_t(int, LDLM_NTHRS_INIT,
3064                                                          ldlm_num_threads));
3065         }
3066
3067         for (i = 0; i < blp->blp_min_threads; i++) {
3068                 rc = ldlm_bl_thread_start(blp, false);
3069                 if (rc < 0)
3070                         GOTO(out, rc);
3071         }
3072
3073 #ifdef HAVE_SERVER_SUPPORT
3074         task = kthread_run(expired_lock_main, NULL, "ldlm_elt");
3075         if (IS_ERR(task)) {
3076                 rc = PTR_ERR(task);
3077                 CERROR("Cannot start ldlm expired-lock thread: %d\n", rc);
3078                 GOTO(out, rc);
3079         }
3080
3081         wait_event(expired_lock_wait_queue,
3082                    expired_lock_thread_state == ELT_READY);
3083 #endif /* HAVE_SERVER_SUPPORT */
3084
3085         rc = ldlm_pools_init();
3086         if (rc) {
3087                 CERROR("Failed to initialize LDLM pools: %d\n", rc);
3088                 GOTO(out, rc);
3089         }
3090
3091         rc = ldlm_reclaim_setup();
3092         if (rc) {
3093                 CERROR("Failed to setup reclaim thread: rc = %d\n", rc);
3094                 GOTO(out, rc);
3095         }
3096         RETURN(0);
3097
3098  out:
3099         ldlm_cleanup();
3100         RETURN(rc);
3101 }
3102
3103 static int ldlm_cleanup(void)
3104 {
3105         ENTRY;
3106
3107         if (!list_empty(ldlm_namespace_list(LDLM_NAMESPACE_SERVER)) ||
3108             !list_empty(ldlm_namespace_list(LDLM_NAMESPACE_CLIENT))) {
3109                 CERROR("ldlm still has namespaces; clean these up first.\n");
3110                 ldlm_dump_all_namespaces(LDLM_NAMESPACE_SERVER, D_DLMTRACE);
3111                 ldlm_dump_all_namespaces(LDLM_NAMESPACE_CLIENT, D_DLMTRACE);
3112                 RETURN(-EBUSY);
3113         }
3114
3115         ldlm_reclaim_cleanup();
3116         ldlm_pools_fini();
3117
3118         if (ldlm_state->ldlm_bl_pool != NULL) {
3119                 struct ldlm_bl_pool *blp = ldlm_state->ldlm_bl_pool;
3120
3121                 while (atomic_read(&blp->blp_num_threads) > 0) {
3122                         struct ldlm_bl_work_item blwi = { .blwi_ns = NULL };
3123
3124                         init_completion(&blp->blp_comp);
3125
3126                         spin_lock(&blp->blp_lock);
3127                         list_add_tail(&blwi.blwi_entry, &blp->blp_list);
3128                         wake_up(&blp->blp_waitq);
3129                         spin_unlock(&blp->blp_lock);
3130
3131                         wait_for_completion(&blp->blp_comp);
3132                 }
3133
3134                 OBD_FREE(blp, sizeof(*blp));
3135         }
3136
3137         if (ldlm_state->ldlm_cb_service != NULL)
3138                 ptlrpc_unregister_service(ldlm_state->ldlm_cb_service);
3139 #ifdef HAVE_SERVER_SUPPORT
3140         if (ldlm_state->ldlm_cancel_service != NULL)
3141                 ptlrpc_unregister_service(ldlm_state->ldlm_cancel_service);
3142 #endif
3143
3144         if (ldlm_ns_kset)
3145                 kset_unregister(ldlm_ns_kset);
3146         if (ldlm_svc_kset)
3147                 kset_unregister(ldlm_svc_kset);
3148         if (ldlm_kobj) {
3149                 sysfs_remove_group(ldlm_kobj, &ldlm_attr_group);
3150                 kobject_put(ldlm_kobj);
3151         }
3152
3153         ldlm_debugfs_cleanup();
3154
3155 #ifdef HAVE_SERVER_SUPPORT
3156         if (expired_lock_thread_state != ELT_STOPPED) {
3157                 expired_lock_thread_state = ELT_TERMINATE;
3158                 wake_up(&expired_lock_wait_queue);
3159                 wait_event(expired_lock_wait_queue,
3160                            expired_lock_thread_state == ELT_STOPPED);
3161         }
3162 #endif
3163
3164         OBD_FREE(ldlm_state, sizeof(*ldlm_state));
3165         ldlm_state = NULL;
3166
3167         RETURN(0);
3168 }
3169
3170 int ldlm_init(void)
3171 {
3172         ldlm_resource_slab = kmem_cache_create("ldlm_resources",
3173                                                sizeof(struct ldlm_resource), 0,
3174                                                SLAB_HWCACHE_ALIGN, NULL);
3175         if (ldlm_resource_slab == NULL)
3176             &n