Whamcloud - gitweb
LU-12616 obclass: fix MDS start/stop race
[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 /*
73  * timeout for initial callback (AST) reply (bz10399)
74  * Due to having to send a 32 bit time value over the
75  * wire return it as time_t instead of time64_t
76  */
77 static inline time_t ldlm_get_rq_timeout(void)
78 {
79         /* Non-AT value */
80         time_t timeout = min(ldlm_timeout, obd_timeout / 3);
81
82         return timeout < 1 ? 1 : timeout;
83 }
84
85 struct ldlm_bl_pool {
86         spinlock_t blp_lock;
87
88         /*
89          * blp_prio_list is used for callbacks that should be handled
90          * as a priority. It is used for LDLM_FL_DISCARD_DATA requests.
91          * see b=13843
92          */
93         struct list_head blp_prio_list;
94
95         /*
96          * blp_list is used for all other callbacks which are likely
97          * to take longer to process.
98          */
99         struct list_head blp_list;
100
101         wait_queue_head_t blp_waitq;
102         struct completion blp_comp;
103         atomic_t blp_num_threads;
104         atomic_t blp_busy_threads;
105         int blp_min_threads;
106         int blp_max_threads;
107 };
108
109 struct ldlm_bl_work_item {
110         struct list_head        blwi_entry;
111         struct ldlm_namespace   *blwi_ns;
112         struct ldlm_lock_desc   blwi_ld;
113         struct ldlm_lock        *blwi_lock;
114         struct list_head        blwi_head;
115         int                     blwi_count;
116         struct completion       blwi_comp;
117         enum ldlm_cancel_flags  blwi_flags;
118         int                     blwi_mem_pressure;
119 };
120
121 #ifdef HAVE_SERVER_SUPPORT
122
123 /**
124  * Protects both waiting_locks_list and expired_lock_thread.
125  */
126 static DEFINE_SPINLOCK(waiting_locks_spinlock); /* BH lock (timer) */
127
128 /**
129  * List for contended locks.
130  *
131  * As soon as a lock is contended, it gets placed on this list and
132  * expected time to get a response is filled in the lock. A special
133  * thread walks the list looking for locks that should be released and
134  * schedules client evictions for those that have not been released in
135  * time.
136  *
137  * All access to it should be under waiting_locks_spinlock.
138  */
139 static LIST_HEAD(waiting_locks_list);
140 static void waiting_locks_callback(TIMER_DATA_TYPE unused);
141 static CFS_DEFINE_TIMER(waiting_locks_timer, waiting_locks_callback, 0, 0);
142
143 enum elt_state {
144         ELT_STOPPED,
145         ELT_READY,
146         ELT_TERMINATE,
147 };
148
149 static DECLARE_WAIT_QUEUE_HEAD(expired_lock_wait_queue);
150 static enum elt_state expired_lock_thread_state = ELT_STOPPED;
151 static int expired_lock_dump;
152 static LIST_HEAD(expired_lock_list);
153
154 static int ldlm_lock_busy(struct ldlm_lock *lock);
155 static int ldlm_add_waiting_lock(struct ldlm_lock *lock, time64_t timeout);
156 static int __ldlm_add_waiting_lock(struct ldlm_lock *lock, time64_t seconds);
157
158 static inline int have_expired_locks(void)
159 {
160         int need_to_run;
161
162         ENTRY;
163         spin_lock_bh(&waiting_locks_spinlock);
164         need_to_run = !list_empty(&expired_lock_list);
165         spin_unlock_bh(&waiting_locks_spinlock);
166
167         RETURN(need_to_run);
168 }
169
170 /**
171  * Check expired lock list for expired locks and time them out.
172  */
173 static int expired_lock_main(void *arg)
174 {
175         struct list_head *expired = &expired_lock_list;
176         struct l_wait_info lwi = { 0 };
177         int do_dump;
178
179         ENTRY;
180
181         expired_lock_thread_state = ELT_READY;
182         wake_up(&expired_lock_wait_queue);
183
184         while (1) {
185                 l_wait_event(expired_lock_wait_queue,
186                              have_expired_locks() ||
187                              expired_lock_thread_state == ELT_TERMINATE,
188                              &lwi);
189
190                 spin_lock_bh(&waiting_locks_spinlock);
191                 if (expired_lock_dump) {
192                         spin_unlock_bh(&waiting_locks_spinlock);
193
194                         /* from waiting_locks_callback, but not in timer */
195                         libcfs_debug_dumplog();
196
197                         spin_lock_bh(&waiting_locks_spinlock);
198                         expired_lock_dump = 0;
199                 }
200
201                 do_dump = 0;
202
203                 while (!list_empty(expired)) {
204                         struct obd_export *export;
205                         struct ldlm_lock *lock;
206
207                         lock = list_entry(expired->next, struct ldlm_lock,
208                                           l_pending_chain);
209                         if ((void *)lock < LP_POISON + PAGE_SIZE &&
210                             (void *)lock >= LP_POISON) {
211                                 spin_unlock_bh(&waiting_locks_spinlock);
212                                 CERROR("free lock on elt list %p\n", lock);
213                                 LBUG();
214                         }
215                         list_del_init(&lock->l_pending_chain);
216                         if ((void *)lock->l_export <
217                              LP_POISON + PAGE_SIZE &&
218                             (void *)lock->l_export >= LP_POISON) {
219                                 CERROR("lock with free export on elt list %p\n",
220                                        lock->l_export);
221                                 lock->l_export = NULL;
222                                 LDLM_ERROR(lock, "free export");
223                                 /*
224                                  * release extra ref grabbed by
225                                  * ldlm_add_waiting_lock() or
226                                  * ldlm_failed_ast()
227                                  */
228                                 LDLM_LOCK_RELEASE(lock);
229                                 continue;
230                         }
231
232                         if (ldlm_is_destroyed(lock)) {
233                                 /*
234                                  * release the lock refcount where
235                                  * waiting_locks_callback() founds
236                                  */
237                                 LDLM_LOCK_RELEASE(lock);
238                                 continue;
239                         }
240                         export = class_export_lock_get(lock->l_export, lock);
241                         spin_unlock_bh(&waiting_locks_spinlock);
242
243                         /* Check if we need to prolong timeout */
244                         if (!OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_TIMEOUT) &&
245                             lock->l_callback_timeout != 0 && /* not AST error */
246                             ldlm_lock_busy(lock)) {
247                                 LDLM_DEBUG(lock, "prolong the busy lock");
248                                 lock_res_and_lock(lock);
249                                 ldlm_add_waiting_lock(lock,
250                                                 ldlm_bl_timeout(lock) >> 1);
251                                 unlock_res_and_lock(lock);
252                         } else {
253                                 spin_lock_bh(&export->exp_bl_list_lock);
254                                 list_del_init(&lock->l_exp_list);
255                                 spin_unlock_bh(&export->exp_bl_list_lock);
256
257                                 LDLM_ERROR(lock,
258                                            "lock callback timer expired after %llds: evicting client at %s ",
259                                            ktime_get_real_seconds() -
260                                            lock->l_blast_sent,
261                                            obd_export_nid2str(export));
262                                 ldlm_lock_to_ns(lock)->ns_timeouts++;
263                                 do_dump++;
264                                 class_fail_export(export);
265                         }
266                         class_export_lock_put(export, lock);
267                         /*
268                          * release extra ref grabbed by ldlm_add_waiting_lock()
269                          * or ldlm_failed_ast()
270                          */
271                         LDLM_LOCK_RELEASE(lock);
272
273                         spin_lock_bh(&waiting_locks_spinlock);
274                 }
275                 spin_unlock_bh(&waiting_locks_spinlock);
276
277                 if (do_dump && obd_dump_on_eviction) {
278                         CERROR("dump the log upon eviction\n");
279                         libcfs_debug_dumplog();
280                 }
281
282                 if (expired_lock_thread_state == ELT_TERMINATE)
283                         break;
284         }
285
286         expired_lock_thread_state = ELT_STOPPED;
287         wake_up(&expired_lock_wait_queue);
288         RETURN(0);
289 }
290
291 /**
292  * Check if there is a request in the export request list
293  * which prevents the lock canceling.
294  */
295 static int ldlm_lock_busy(struct ldlm_lock *lock)
296 {
297         struct ptlrpc_request *req;
298         int match = 0;
299
300         ENTRY;
301
302         if (lock->l_export == NULL)
303                 return 0;
304
305         spin_lock(&lock->l_export->exp_rpc_lock);
306         list_for_each_entry(req, &lock->l_export->exp_hp_rpcs,
307                                 rq_exp_list) {
308                 if (req->rq_ops->hpreq_lock_match) {
309                         match = req->rq_ops->hpreq_lock_match(req, lock);
310                         if (match)
311                                 break;
312                 }
313         }
314         spin_unlock(&lock->l_export->exp_rpc_lock);
315         RETURN(match);
316 }
317
318 /* This is called from within a timer interrupt and cannot schedule */
319 static void waiting_locks_callback(TIMER_DATA_TYPE unused)
320 {
321         struct ldlm_lock *lock;
322         int need_dump = 0;
323
324         spin_lock_bh(&waiting_locks_spinlock);
325         while (!list_empty(&waiting_locks_list)) {
326                 lock = list_entry(waiting_locks_list.next, struct ldlm_lock,
327                                   l_pending_chain);
328                 if (lock->l_callback_timeout > ktime_get_seconds() ||
329                     lock->l_req_mode == LCK_GROUP)
330                         break;
331
332                 /*
333                  * no needs to take an extra ref on the lock since it was in
334                  * the waiting_locks_list and ldlm_add_waiting_lock()
335                  * already grabbed a ref
336                  */
337                 list_del(&lock->l_pending_chain);
338                 list_add(&lock->l_pending_chain, &expired_lock_list);
339                 need_dump = 1;
340         }
341
342         if (!list_empty(&expired_lock_list)) {
343                 if (obd_dump_on_timeout && need_dump)
344                         expired_lock_dump = __LINE__;
345
346                 wake_up(&expired_lock_wait_queue);
347         }
348
349         /*
350          * Make sure the timer will fire again if we have any locks
351          * left.
352          */
353         if (!list_empty(&waiting_locks_list)) {
354                 unsigned long timeout_jiffies;
355
356                 lock = list_entry(waiting_locks_list.next, struct ldlm_lock,
357                                   l_pending_chain);
358                 timeout_jiffies = cfs_time_seconds(lock->l_callback_timeout);
359                 mod_timer(&waiting_locks_timer, timeout_jiffies);
360         }
361         spin_unlock_bh(&waiting_locks_spinlock);
362 }
363
364 /**
365  * Add lock to the list of contended locks.
366  *
367  * Indicate that we're waiting for a client to call us back cancelling a given
368  * lock.  We add it to the pending-callback chain, and schedule the lock-timeout
369  * timer to fire appropriately.  (We round up to the next second, to avoid
370  * floods of timer firings during periods of high lock contention and traffic).
371  * As done by ldlm_add_waiting_lock(), the caller must grab a lock reference
372  * if it has been added to the waiting list (1 is returned).
373  *
374  * Called with the namespace lock held.
375  */
376 static int __ldlm_add_waiting_lock(struct ldlm_lock *lock, time64_t seconds)
377 {
378         unsigned long timeout_jiffies;
379         time64_t timeout;
380
381         if (!list_empty(&lock->l_pending_chain))
382                 return 0;
383
384         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_NOTIMEOUT) ||
385             OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_TIMEOUT))
386                 seconds = 1;
387
388         timeout = ktime_get_seconds() + seconds;
389         if (likely(timeout > lock->l_callback_timeout))
390                 lock->l_callback_timeout = timeout;
391
392         timeout_jiffies = cfs_time_seconds(lock->l_callback_timeout);
393
394         if (time_before(timeout_jiffies, waiting_locks_timer.expires) ||
395             !timer_pending(&waiting_locks_timer))
396                 mod_timer(&waiting_locks_timer, timeout_jiffies);
397
398         /*
399          * if the new lock has a shorter timeout than something earlier on
400          * the list, we'll wait the longer amount of time; no big deal.
401          */
402         /* FIFO */
403         list_add_tail(&lock->l_pending_chain, &waiting_locks_list);
404         return 1;
405 }
406
407 static void ldlm_add_blocked_lock(struct ldlm_lock *lock)
408 {
409         spin_lock_bh(&lock->l_export->exp_bl_list_lock);
410         if (list_empty(&lock->l_exp_list)) {
411                 if (!ldlm_is_granted(lock))
412                         list_add_tail(&lock->l_exp_list,
413                                       &lock->l_export->exp_bl_list);
414                 else
415                         list_add(&lock->l_exp_list,
416                                  &lock->l_export->exp_bl_list);
417         }
418         spin_unlock_bh(&lock->l_export->exp_bl_list_lock);
419
420         /*
421          * A blocked lock is added. Adjust the position in
422          * the stale list if the export is in the list.
423          * If export is stale and not in the list - it is being
424          * processed and will be placed on the right position
425          * on obd_stale_export_put().
426          */
427         if (!list_empty(&lock->l_export->exp_stale_list))
428                 obd_stale_export_adjust(lock->l_export);
429 }
430
431 static int ldlm_add_waiting_lock(struct ldlm_lock *lock, time64_t timeout)
432 {
433         int ret;
434
435         /* NB: must be called with hold of lock_res_and_lock() */
436         LASSERT(ldlm_is_res_locked(lock));
437         LASSERT(!ldlm_is_cancel_on_block(lock));
438
439         /*
440          * Do not put cross-MDT lock in the waiting list, since we
441          * will not evict it due to timeout for now
442          */
443         if (lock->l_export != NULL &&
444             (exp_connect_flags(lock->l_export) & OBD_CONNECT_MDS_MDS))
445                 return 0;
446
447         spin_lock_bh(&waiting_locks_spinlock);
448         if (ldlm_is_cancel(lock)) {
449                 spin_unlock_bh(&waiting_locks_spinlock);
450                 return 0;
451         }
452
453         if (ldlm_is_destroyed(lock)) {
454                 static time64_t next;
455
456                 spin_unlock_bh(&waiting_locks_spinlock);
457                 LDLM_ERROR(lock, "not waiting on destroyed lock (b=5653)");
458                 if (ktime_get_seconds() > next) {
459                         next = ktime_get_seconds() + 14400;
460                         libcfs_debug_dumpstack(NULL);
461                 }
462                 return 0;
463         }
464
465         ldlm_set_waited(lock);
466         lock->l_blast_sent = ktime_get_real_seconds();
467         ret = __ldlm_add_waiting_lock(lock, timeout);
468         if (ret) {
469                 /*
470                  * grab ref on the lock if it has been added to the
471                  * waiting list
472                  */
473                 LDLM_LOCK_GET(lock);
474         }
475         spin_unlock_bh(&waiting_locks_spinlock);
476
477         if (ret)
478                 ldlm_add_blocked_lock(lock);
479
480         LDLM_DEBUG(lock, "%sadding to wait list(timeout: %lld, AT: %s)",
481                    ret == 0 ? "not re-" : "", timeout,
482                    AT_OFF ? "off" : "on");
483         return ret;
484 }
485
486 /**
487  * Remove a lock from the pending list, likely because it had its cancellation
488  * callback arrive without incident.  This adjusts the lock-timeout timer if
489  * needed.  Returns 0 if the lock wasn't pending after all, 1 if it was.
490  * As done by ldlm_del_waiting_lock(), the caller must release the lock
491  * reference when the lock is removed from any list (1 is returned).
492  *
493  * Called with namespace lock held.
494  */
495 static int __ldlm_del_waiting_lock(struct ldlm_lock *lock)
496 {
497         struct list_head *list_next;
498
499         if (list_empty(&lock->l_pending_chain))
500                 return 0;
501
502         list_next = lock->l_pending_chain.next;
503         if (lock->l_pending_chain.prev == &waiting_locks_list) {
504                 /* Removing the head of the list, adjust timer. */
505                 if (list_next == &waiting_locks_list) {
506                         /* No more, just cancel. */
507                         del_timer(&waiting_locks_timer);
508                 } else {
509                         struct ldlm_lock *next;
510
511                         next = list_entry(list_next, struct ldlm_lock,
512                                           l_pending_chain);
513                         mod_timer(&waiting_locks_timer,
514                                   cfs_time_seconds(next->l_callback_timeout));
515                 }
516         }
517         list_del_init(&lock->l_pending_chain);
518
519         return 1;
520 }
521
522 int ldlm_del_waiting_lock(struct ldlm_lock *lock)
523 {
524         int ret;
525
526         if (lock->l_export == NULL) {
527                 /* We don't have a "waiting locks list" on clients. */
528                 CDEBUG(D_DLMTRACE, "Client lock %p : no-op\n", lock);
529                 return 0;
530         }
531
532         spin_lock_bh(&waiting_locks_spinlock);
533         ret = __ldlm_del_waiting_lock(lock);
534         ldlm_clear_waited(lock);
535         spin_unlock_bh(&waiting_locks_spinlock);
536
537         /* remove the lock out of export blocking list */
538         spin_lock_bh(&lock->l_export->exp_bl_list_lock);
539         list_del_init(&lock->l_exp_list);
540         spin_unlock_bh(&lock->l_export->exp_bl_list_lock);
541
542         if (ret) {
543                 /*
544                  * release lock ref if it has indeed been removed
545                  * from a list
546                  */
547                 LDLM_LOCK_RELEASE(lock);
548         }
549
550         LDLM_DEBUG(lock, "%s", ret == 0 ? "wasn't waiting" : "removed");
551         return ret;
552 }
553
554 /**
555  * Prolong the contended lock waiting time.
556  *
557  * Called with namespace lock held.
558  */
559 int ldlm_refresh_waiting_lock(struct ldlm_lock *lock, time64_t timeout)
560 {
561         if (lock->l_export == NULL) {
562                 /* We don't have a "waiting locks list" on clients. */
563                 LDLM_DEBUG(lock, "client lock: no-op");
564                 return 0;
565         }
566
567         if (exp_connect_flags(lock->l_export) & OBD_CONNECT_MDS_MDS) {
568                 /* We don't have a "waiting locks list" on OSP. */
569                 LDLM_DEBUG(lock, "MDS-MDS lock: no-op");
570                 return 0;
571         }
572
573         spin_lock_bh(&waiting_locks_spinlock);
574
575         if (list_empty(&lock->l_pending_chain)) {
576                 spin_unlock_bh(&waiting_locks_spinlock);
577                 LDLM_DEBUG(lock, "wasn't waiting");
578                 return 0;
579         }
580
581         /*
582          * we remove/add the lock to the waiting list, so no needs to
583          * release/take a lock reference
584          */
585         __ldlm_del_waiting_lock(lock);
586         __ldlm_add_waiting_lock(lock, timeout);
587         spin_unlock_bh(&waiting_locks_spinlock);
588
589         LDLM_DEBUG(lock, "refreshed");
590         return 1;
591 }
592 EXPORT_SYMBOL(ldlm_refresh_waiting_lock);
593
594 #else /* HAVE_SERVER_SUPPORT */
595
596 int ldlm_del_waiting_lock(struct ldlm_lock *lock)
597 {
598         RETURN(0);
599 }
600
601 int ldlm_refresh_waiting_lock(struct ldlm_lock *lock, time64_t timeout)
602 {
603         RETURN(0);
604 }
605
606 #endif /* !HAVE_SERVER_SUPPORT */
607
608 #ifdef HAVE_SERVER_SUPPORT
609
610 /**
611  * Calculate the per-export Blocking timeout (covering BL AST, data flush,
612  * lock cancel, and their replies). Used for lock callback timeout and AST
613  * re-send period.
614  *
615  * \param[in] lock        lock which is getting the blocking callback
616  *
617  * \retval            timeout in seconds to wait for the client reply
618  */
619 time64_t ldlm_bl_timeout(struct ldlm_lock *lock)
620 {
621         time64_t timeout;
622
623         if (AT_OFF)
624                 return obd_timeout / 2;
625
626         /*
627          * Since these are non-updating timeouts, we should be conservative.
628          * Take more than usually, 150%
629          * It would be nice to have some kind of "early reply" mechanism for
630          * lock callbacks too...
631          */
632         timeout = at_get(&lock->l_export->exp_bl_lock_at);
633         return max(timeout + (timeout >> 1), (time64_t)ldlm_enqueue_min);
634 }
635 EXPORT_SYMBOL(ldlm_bl_timeout);
636
637 /**
638  * Perform lock cleanup if AST sending failed.
639  */
640 static void ldlm_failed_ast(struct ldlm_lock *lock, int rc,
641                             const char *ast_type)
642 {
643         LCONSOLE_ERROR_MSG(0x138,
644                            "%s: A client on nid %s was evicted due to a lock %s callback time out: rc %d\n",
645                            lock->l_export->exp_obd->obd_name,
646                            obd_export_nid2str(lock->l_export), ast_type, rc);
647
648         if (obd_dump_on_timeout)
649                 libcfs_debug_dumplog();
650         spin_lock_bh(&waiting_locks_spinlock);
651         if (__ldlm_del_waiting_lock(lock) == 0)
652                 /*
653                  * the lock was not in any list, grab an extra ref before adding
654                  * the lock to the expired list
655                  */
656                 LDLM_LOCK_GET(lock);
657         lock->l_callback_timeout = 0; /* differentiate it from expired locks */
658         list_add(&lock->l_pending_chain, &expired_lock_list);
659         wake_up(&expired_lock_wait_queue);
660         spin_unlock_bh(&waiting_locks_spinlock);
661 }
662
663 /**
664  * Perform lock cleanup if AST reply came with error.
665  */
666 static int ldlm_handle_ast_error(struct ldlm_lock *lock,
667                                  struct ptlrpc_request *req, int rc,
668                                  const char *ast_type)
669 {
670         struct lnet_process_id peer = req->rq_import->imp_connection->c_peer;
671
672         if (!req->rq_replied || (rc && rc != -EINVAL)) {
673                 if (ldlm_is_cancel(lock)) {
674                         LDLM_DEBUG(lock,
675                                    "%s AST (req@%p x%llu) timeout from nid %s, but cancel was received (AST reply lost?)",
676                                    ast_type, req, req->rq_xid,
677                                    libcfs_nid2str(peer.nid));
678                         ldlm_lock_cancel(lock);
679                         rc = -ERESTART;
680                 } else if (rc == -ENODEV || rc == -ESHUTDOWN ||
681                            (rc == -EIO &&
682                             req->rq_import->imp_state == LUSTRE_IMP_CLOSED)) {
683                         /*
684                          * Upon umount process the AST fails because cannot be
685                          * sent. This shouldn't lead to the client eviction.
686                          * -ENODEV error is returned by ptl_send_rpc() for
687                          *  new request in such import.
688                          * -SHUTDOWN is returned by ptlrpc_import_delay_req()
689                          *  if imp_invalid is set or obd_no_recov.
690                          * Meanwhile there is also check for LUSTRE_IMP_CLOSED
691                          * in ptlrpc_import_delay_req() as well with -EIO code.
692                          * In all such cases errors are ignored.
693                          */
694                         LDLM_DEBUG(lock,
695                                    "%s AST can't be sent due to a server %s failure or umount process: rc = %d\n",
696                                     ast_type,
697                                      req->rq_import->imp_obd->obd_name, rc);
698                 } else {
699                         LDLM_ERROR(lock,
700                                    "client (nid %s) %s %s AST (req@%p x%llu status %d rc %d), evict it",
701                                    libcfs_nid2str(peer.nid),
702                                    req->rq_replied ? "returned error from" :
703                                    "failed to reply to",
704                                    ast_type, req, req->rq_xid,
705                                    (req->rq_repmsg != NULL) ?
706                                    lustre_msg_get_status(req->rq_repmsg) : 0,
707                                    rc);
708                         ldlm_failed_ast(lock, rc, ast_type);
709                 }
710                 return rc;
711         }
712
713         if (rc == -EINVAL) {
714                 struct ldlm_resource *res = lock->l_resource;
715
716                 LDLM_DEBUG(lock,
717                            "client (nid %s) returned %d from %s AST (req@%p x%llu) - normal race",
718                            libcfs_nid2str(peer.nid),
719                            req->rq_repmsg ?
720                            lustre_msg_get_status(req->rq_repmsg) : -1,
721                            ast_type, req, req->rq_xid);
722                 if (res) {
723                         /*
724                          * update lvbo to return proper attributes.
725                          * see b=23174
726                          */
727                         ldlm_resource_getref(res);
728                         ldlm_lvbo_update(res, lock, NULL, 1);
729                         ldlm_resource_putref(res);
730                 }
731                 ldlm_lock_cancel(lock);
732                 rc = -ERESTART;
733         }
734
735         return rc;
736 }
737
738 static int ldlm_cb_interpret(const struct lu_env *env,
739                              struct ptlrpc_request *req, void *args, int rc)
740 {
741         struct ldlm_cb_async_args *ca = args;
742         struct ldlm_lock *lock = ca->ca_lock;
743         struct ldlm_cb_set_arg *arg  = ca->ca_set_arg;
744
745         ENTRY;
746
747         LASSERT(lock != NULL);
748
749         switch (arg->type) {
750         case LDLM_GL_CALLBACK:
751                 /*
752                  * Update the LVB from disk if the AST failed
753                  * (this is a legal race)
754                  *
755                  * - Glimpse callback of local lock just returns
756                  *   -ELDLM_NO_LOCK_DATA.
757                  * - Glimpse callback of remote lock might return
758                  *   -ELDLM_NO_LOCK_DATA when inode is cleared. LU-274
759                  */
760                 if (unlikely(arg->gl_interpret_reply)) {
761                         rc = arg->gl_interpret_reply(NULL, req, args, rc);
762                 } else if (rc == -ELDLM_NO_LOCK_DATA) {
763                         LDLM_DEBUG(lock,
764                                    "lost race - client has a lock but no inode");
765                         ldlm_lvbo_update(lock->l_resource, lock, NULL, 1);
766                 } else if (rc != 0) {
767                         rc = ldlm_handle_ast_error(lock, req, rc, "glimpse");
768                 } else {
769                         rc = ldlm_lvbo_update(lock->l_resource,
770                                               lock, req, 1);
771                 }
772                 break;
773         case LDLM_BL_CALLBACK:
774                 if (rc != 0)
775                         rc = ldlm_handle_ast_error(lock, req, rc, "blocking");
776                 break;
777         case LDLM_CP_CALLBACK:
778                 if (rc != 0)
779                         rc = ldlm_handle_ast_error(lock, req, rc, "completion");
780                 break;
781         default:
782                 LDLM_ERROR(lock, "invalid opcode for lock callback %d",
783                            arg->type);
784                 LBUG();
785         }
786
787         /* release extra reference taken in ldlm_ast_fini() */
788         LDLM_LOCK_RELEASE(lock);
789
790         if (rc == -ERESTART)
791                 atomic_inc(&arg->restart);
792
793         RETURN(0);
794 }
795
796 static void ldlm_update_resend(struct ptlrpc_request *req, void *data)
797 {
798         struct ldlm_cb_async_args *ca = data;
799         struct ldlm_lock *lock = ca->ca_lock;
800
801         ldlm_refresh_waiting_lock(lock, ldlm_bl_timeout(lock));
802 }
803
804 static inline int ldlm_ast_fini(struct ptlrpc_request *req,
805                                 struct ldlm_cb_set_arg *arg,
806                                 struct ldlm_lock *lock,
807                                 int instant_cancel)
808 {
809         int rc = 0;
810
811         ENTRY;
812
813         if (unlikely(instant_cancel)) {
814                 rc = ptl_send_rpc(req, 1);
815                 ptlrpc_req_finished(req);
816                 if (rc == 0)
817                         atomic_inc(&arg->restart);
818         } else {
819                 LDLM_LOCK_GET(lock);
820                 ptlrpc_set_add_req(arg->set, req);
821         }
822
823         RETURN(rc);
824 }
825
826 /**
827  * Check if there are requests in the export request list which prevent
828  * the lock canceling and make these requests high priority ones.
829  */
830 static void ldlm_lock_reorder_req(struct ldlm_lock *lock)
831 {
832         struct ptlrpc_request *req;
833
834         ENTRY;
835
836         if (lock->l_export == NULL) {
837                 LDLM_DEBUG(lock, "client lock: no-op");
838                 RETURN_EXIT;
839         }
840
841         spin_lock(&lock->l_export->exp_rpc_lock);
842         list_for_each_entry(req, &lock->l_export->exp_hp_rpcs,
843                             rq_exp_list) {
844                 /*
845                  * Do not process requests that were not yet added to there
846                  * incoming queue or were already removed from there for
847                  * processing. We evaluate ptlrpc_nrs_req_can_move() without
848                  * holding svcpt->scp_req_lock, and then redo the check with
849                  * the lock held once we need to obtain a reliable result.
850                  */
851                 if (ptlrpc_nrs_req_can_move(req) &&
852                     req->rq_ops->hpreq_lock_match &&
853                     req->rq_ops->hpreq_lock_match(req, lock))
854                         ptlrpc_nrs_req_hp_move(req);
855         }
856         spin_unlock(&lock->l_export->exp_rpc_lock);
857         EXIT;
858 }
859
860 /**
861  * ->l_blocking_ast() method for server-side locks. This is invoked when newly
862  * enqueued server lock conflicts with given one.
863  *
864  * Sends blocking AST RPC to the client owning that lock; arms timeout timer
865  * to wait for client response.
866  */
867 int ldlm_server_blocking_ast(struct ldlm_lock *lock,
868                              struct ldlm_lock_desc *desc,
869                              void *data, int flag)
870 {
871         struct ldlm_cb_async_args *ca;
872         struct ldlm_cb_set_arg *arg = data;
873         struct ldlm_request *body;
874         struct ptlrpc_request  *req;
875         int instant_cancel = 0;
876         int rc = 0;
877
878         ENTRY;
879
880         if (flag == LDLM_CB_CANCELING)
881                 /* Don't need to do anything here. */
882                 RETURN(0);
883
884         if (OBD_FAIL_PRECHECK(OBD_FAIL_LDLM_SRV_BL_AST)) {
885                 LDLM_DEBUG(lock, "dropping BL AST");
886                 RETURN(0);
887         }
888
889         LASSERT(lock);
890         LASSERT(data != NULL);
891         if (lock->l_export->exp_obd->obd_recovering != 0)
892                 LDLM_ERROR(lock, "BUG 6063: lock collide during recovery");
893
894         ldlm_lock_reorder_req(lock);
895
896         req = ptlrpc_request_alloc_pack(lock->l_export->exp_imp_reverse,
897                                         &RQF_LDLM_BL_CALLBACK,
898                                         LUSTRE_DLM_VERSION, LDLM_BL_CALLBACK);
899         if (req == NULL)
900                 RETURN(-ENOMEM);
901
902         CLASSERT(sizeof(*ca) <= sizeof(req->rq_async_args));
903         ca = ptlrpc_req_async_args(req);
904         ca->ca_set_arg = arg;
905         ca->ca_lock = lock;
906
907         req->rq_interpret_reply = ldlm_cb_interpret;
908
909         lock_res_and_lock(lock);
910         if (ldlm_is_destroyed(lock)) {
911                 /* What's the point? */
912                 unlock_res_and_lock(lock);
913                 ptlrpc_req_finished(req);
914                 RETURN(0);
915         }
916
917         if (!ldlm_is_granted(lock)) {
918                 /*
919                  * this blocking AST will be communicated as part of the
920                  * completion AST instead
921                  */
922                 ldlm_add_blocked_lock(lock);
923                 ldlm_set_waited(lock);
924                 unlock_res_and_lock(lock);
925
926                 ptlrpc_req_finished(req);
927                 LDLM_DEBUG(lock, "lock not granted, not sending blocking AST");
928                 RETURN(0);
929         }
930
931         if (ldlm_is_cancel_on_block(lock))
932                 instant_cancel = 1;
933
934         body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
935         body->lock_handle[0] = lock->l_remote_handle;
936         body->lock_desc = *desc;
937         body->lock_flags |= ldlm_flags_to_wire(lock->l_flags & LDLM_FL_AST_MASK);
938
939         LDLM_DEBUG(lock, "server preparing blocking AST");
940
941         ptlrpc_request_set_replen(req);
942         ldlm_set_cbpending(lock);
943         if (instant_cancel) {
944                 unlock_res_and_lock(lock);
945                 ldlm_lock_cancel(lock);
946
947                 req->rq_no_resend = 1;
948         } else {
949                 LASSERT(ldlm_is_granted(lock));
950                 ldlm_add_waiting_lock(lock, ldlm_bl_timeout(lock));
951                 unlock_res_and_lock(lock);
952
953                 /* Do not resend after lock callback timeout */
954                 req->rq_delay_limit = ldlm_bl_timeout(lock);
955                 req->rq_resend_cb = ldlm_update_resend;
956         }
957
958         req->rq_send_state = LUSTRE_IMP_FULL;
959         /* ptlrpc_request_alloc_pack already set timeout */
960         if (AT_OFF)
961                 req->rq_timeout = ldlm_get_rq_timeout();
962
963         if (lock->l_export && lock->l_export->exp_nid_stats &&
964             lock->l_export->exp_nid_stats->nid_ldlm_stats)
965                 lprocfs_counter_incr(lock->l_export->exp_nid_stats->nid_ldlm_stats,
966                                      LDLM_BL_CALLBACK - LDLM_FIRST_OPC);
967
968         rc = ldlm_ast_fini(req, arg, lock, instant_cancel);
969
970         RETURN(rc);
971 }
972
973 /**
974  * ->l_completion_ast callback for a remote lock in server namespace.
975  *
976  *  Sends AST to the client notifying it of lock granting.  If initial
977  *  lock response was not sent yet, instead of sending another RPC, just
978  *  mark the lock as granted and client will understand
979  */
980 int ldlm_server_completion_ast(struct ldlm_lock *lock, __u64 flags, void *data)
981 {
982         struct ldlm_cb_set_arg *arg = data;
983         struct ldlm_request *body;
984         struct ptlrpc_request *req;
985         struct ldlm_cb_async_args *ca;
986         int instant_cancel = 0;
987         int rc = 0;
988         int lvb_len;
989
990         ENTRY;
991
992         LASSERT(lock != NULL);
993         LASSERT(data != NULL);
994
995         if (OBD_FAIL_PRECHECK(OBD_FAIL_LDLM_SRV_CP_AST)) {
996                 LDLM_DEBUG(lock, "dropping CP AST");
997                 RETURN(0);
998         }
999
1000         req = ptlrpc_request_alloc(lock->l_export->exp_imp_reverse,
1001                                    &RQF_LDLM_CP_CALLBACK);
1002         if (req == NULL)
1003                 RETURN(-ENOMEM);
1004
1005         /* server namespace, doesn't need lock */
1006         lvb_len = ldlm_lvbo_size(lock);
1007         /*
1008          * LU-3124 & LU-2187: to not return layout in completion AST because
1009          * it may deadlock for LU-2187, or client may not have enough space
1010          * for large layout. The layout will be returned to client with an
1011          * extra RPC to fetch xattr.lov
1012          */
1013         if (ldlm_has_layout(lock))
1014                 lvb_len = 0;
1015
1016         req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_CLIENT, lvb_len);
1017         rc = ptlrpc_request_pack(req, LUSTRE_DLM_VERSION, LDLM_CP_CALLBACK);
1018         if (rc) {
1019                 ptlrpc_request_free(req);
1020                 RETURN(rc);
1021         }
1022
1023         CLASSERT(sizeof(*ca) <= sizeof(req->rq_async_args));
1024         ca = ptlrpc_req_async_args(req);
1025         ca->ca_set_arg = arg;
1026         ca->ca_lock = lock;
1027
1028         req->rq_interpret_reply = ldlm_cb_interpret;
1029         body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1030
1031         body->lock_handle[0] = lock->l_remote_handle;
1032         body->lock_flags = ldlm_flags_to_wire(flags);
1033         ldlm_lock2desc(lock, &body->lock_desc);
1034         if (lvb_len > 0) {
1035                 void *lvb = req_capsule_client_get(&req->rq_pill, &RMF_DLM_LVB);
1036                 lvb_len = ldlm_lvbo_fill(lock, lvb, &lvb_len);
1037                 if (lvb_len < 0) {
1038                         /*
1039                          * We still need to send the RPC to wake up the blocked
1040                          * enqueue thread on the client.
1041                          *
1042                          * Consider old client, there is no better way to notify
1043                          * the failure, just zero-sized the LVB, then the client
1044                          * will fail out as "-EPROTO".
1045                          */
1046                         req_capsule_shrink(&req->rq_pill, &RMF_DLM_LVB, 0,
1047                                            RCL_CLIENT);
1048                         instant_cancel = 1;
1049                 } else {
1050                         req_capsule_shrink(&req->rq_pill, &RMF_DLM_LVB, lvb_len,
1051                                            RCL_CLIENT);
1052                 }
1053         }
1054
1055         LDLM_DEBUG(lock, "server preparing completion AST");
1056
1057         ptlrpc_request_set_replen(req);
1058
1059         req->rq_send_state = LUSTRE_IMP_FULL;
1060         /* ptlrpc_request_pack already set timeout */
1061         if (AT_OFF)
1062                 req->rq_timeout = ldlm_get_rq_timeout();
1063
1064         /* We only send real blocking ASTs after the lock is granted */
1065         lock_res_and_lock(lock);
1066         if (ldlm_is_ast_sent(lock)) {
1067                 body->lock_flags |= ldlm_flags_to_wire(LDLM_FL_AST_SENT);
1068                 /* Copy AST flags like LDLM_FL_DISCARD_DATA. */
1069                 body->lock_flags |= ldlm_flags_to_wire(lock->l_flags &
1070                                                        LDLM_FL_AST_MASK);
1071
1072                 /*
1073                  * We might get here prior to ldlm_handle_enqueue setting
1074                  * LDLM_FL_CANCEL_ON_BLOCK flag. Then we will put this lock
1075                  * into waiting list, but this is safe and similar code in
1076                  * ldlm_handle_enqueue will call ldlm_lock_cancel() still,
1077                  * that would not only cancel the lock, but will also remove
1078                  * it from waiting list
1079                  */
1080                 if (ldlm_is_cancel_on_block(lock)) {
1081                         unlock_res_and_lock(lock);
1082                         ldlm_lock_cancel(lock);
1083
1084                         instant_cancel = 1;
1085                         req->rq_no_resend = 1;
1086
1087                         lock_res_and_lock(lock);
1088                 } else {
1089                         /* start the lock-timeout clock */
1090                         ldlm_add_waiting_lock(lock, ldlm_bl_timeout(lock));
1091                         /* Do not resend after lock callback timeout */
1092                         req->rq_delay_limit = ldlm_bl_timeout(lock);
1093                         req->rq_resend_cb = ldlm_update_resend;
1094                 }
1095         }
1096         unlock_res_and_lock(lock);
1097
1098         if (lock->l_export && lock->l_export->exp_nid_stats &&
1099             lock->l_export->exp_nid_stats->nid_ldlm_stats)
1100                 lprocfs_counter_incr(lock->l_export->exp_nid_stats->nid_ldlm_stats,
1101                                      LDLM_CP_CALLBACK - LDLM_FIRST_OPC);
1102
1103         rc = ldlm_ast_fini(req, arg, lock, instant_cancel);
1104
1105         RETURN(lvb_len < 0 ? lvb_len : rc);
1106 }
1107
1108 /**
1109  * Server side ->l_glimpse_ast handler for client locks.
1110  *
1111  * Sends glimpse AST to the client and waits for reply. Then updates
1112  * lvbo with the result.
1113  */
1114 int ldlm_server_glimpse_ast(struct ldlm_lock *lock, void *data)
1115 {
1116         struct ldlm_cb_set_arg *arg = data;
1117         struct ldlm_request *body;
1118         struct ptlrpc_request *req;
1119         struct ldlm_cb_async_args *ca;
1120         int rc;
1121         struct req_format *req_fmt;
1122
1123         ENTRY;
1124
1125         LASSERT(lock != NULL);
1126
1127         if (arg->gl_desc != NULL)
1128                 /* There is a glimpse descriptor to pack */
1129                 req_fmt = &RQF_LDLM_GL_CALLBACK_DESC;
1130         else
1131                 req_fmt = &RQF_LDLM_GL_CALLBACK;
1132
1133         req = ptlrpc_request_alloc_pack(lock->l_export->exp_imp_reverse,
1134                                         req_fmt, LUSTRE_DLM_VERSION,
1135                                         LDLM_GL_CALLBACK);
1136
1137         if (req == NULL)
1138                 RETURN(-ENOMEM);
1139
1140         if (arg->gl_desc != NULL) {
1141                 /* copy the GL descriptor */
1142                 union ldlm_gl_desc      *desc;
1143
1144                 desc = req_capsule_client_get(&req->rq_pill, &RMF_DLM_GL_DESC);
1145                 *desc = *arg->gl_desc;
1146         }
1147
1148         body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1149         body->lock_handle[0] = lock->l_remote_handle;
1150         ldlm_lock2desc(lock, &body->lock_desc);
1151
1152         CLASSERT(sizeof(*ca) <= sizeof(req->rq_async_args));
1153         ca = ptlrpc_req_async_args(req);
1154         ca->ca_set_arg = arg;
1155         ca->ca_lock = lock;
1156
1157         /* server namespace, doesn't need lock */
1158         req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
1159                              ldlm_lvbo_size(lock));
1160         ptlrpc_request_set_replen(req);
1161
1162         req->rq_send_state = LUSTRE_IMP_FULL;
1163         /* ptlrpc_request_alloc_pack already set timeout */
1164         if (AT_OFF)
1165                 req->rq_timeout = ldlm_get_rq_timeout();
1166
1167         req->rq_interpret_reply = ldlm_cb_interpret;
1168
1169         if (lock->l_export && lock->l_export->exp_nid_stats) {
1170                 struct nid_stat *nid_stats = lock->l_export->exp_nid_stats;
1171
1172                 lprocfs_counter_incr(nid_stats->nid_ldlm_stats,
1173                                      LDLM_GL_CALLBACK - LDLM_FIRST_OPC);
1174         }
1175
1176         rc = ldlm_ast_fini(req, arg, lock, 0);
1177
1178         RETURN(rc);
1179 }
1180 EXPORT_SYMBOL(ldlm_server_glimpse_ast);
1181
1182 int ldlm_glimpse_locks(struct ldlm_resource *res,
1183                        struct list_head *gl_work_list)
1184 {
1185         int rc;
1186
1187         ENTRY;
1188
1189         rc = ldlm_run_ast_work(ldlm_res_to_ns(res), gl_work_list,
1190                                LDLM_WORK_GL_AST);
1191         if (rc == -ERESTART)
1192                 ldlm_reprocess_all(res, NULL);
1193
1194         RETURN(rc);
1195 }
1196 EXPORT_SYMBOL(ldlm_glimpse_locks);
1197
1198 /* return LDLM lock associated with a lock callback request */
1199 struct ldlm_lock *ldlm_request_lock(struct ptlrpc_request *req)
1200 {
1201         struct ldlm_cb_async_args *ca;
1202         struct ldlm_lock *lock;
1203
1204         ENTRY;
1205
1206         ca = ptlrpc_req_async_args(req);
1207         lock = ca->ca_lock;
1208         if (lock == NULL)
1209                 RETURN(ERR_PTR(-EFAULT));
1210
1211         RETURN(lock);
1212 }
1213 EXPORT_SYMBOL(ldlm_request_lock);
1214
1215 /**
1216  * Main server-side entry point into LDLM for enqueue. This is called by ptlrpc
1217  * service threads to carry out client lock enqueueing requests.
1218  */
1219 int ldlm_handle_enqueue0(struct ldlm_namespace *ns,
1220                          struct ptlrpc_request *req,
1221                          const struct ldlm_request *dlm_req,
1222                          const struct ldlm_callback_suite *cbs)
1223 {
1224         struct ldlm_reply *dlm_rep;
1225         __u64 flags;
1226         enum ldlm_error err = ELDLM_OK;
1227         struct ldlm_lock *lock = NULL;
1228         void *cookie = NULL;
1229         int rc = 0;
1230         struct ldlm_resource *res = NULL;
1231         const struct lu_env *env = req->rq_svc_thread->t_env;
1232
1233         ENTRY;
1234
1235         LDLM_DEBUG_NOLOCK("server-side enqueue handler START");
1236
1237         ldlm_request_cancel(req, dlm_req, LDLM_ENQUEUE_CANCEL_OFF, LATF_SKIP);
1238         flags = ldlm_flags_from_wire(dlm_req->lock_flags);
1239
1240         LASSERT(req->rq_export);
1241
1242         /* for intent enqueue the stat will be updated inside intent policy */
1243         if (ptlrpc_req2svc(req)->srv_stats != NULL &&
1244             !(dlm_req->lock_flags & LDLM_FL_HAS_INTENT))
1245                 ldlm_svc_get_eopc(dlm_req, ptlrpc_req2svc(req)->srv_stats);
1246
1247         if (req->rq_export && req->rq_export->exp_nid_stats &&
1248             req->rq_export->exp_nid_stats->nid_ldlm_stats)
1249                 lprocfs_counter_incr(req->rq_export->exp_nid_stats->nid_ldlm_stats,
1250                                      LDLM_ENQUEUE - LDLM_FIRST_OPC);
1251
1252         if (unlikely(dlm_req->lock_desc.l_resource.lr_type < LDLM_MIN_TYPE ||
1253                      dlm_req->lock_desc.l_resource.lr_type >= LDLM_MAX_TYPE)) {
1254                 DEBUG_REQ(D_ERROR, req, "invalid lock request type %d",
1255                           dlm_req->lock_desc.l_resource.lr_type);
1256                 GOTO(out, rc = -EFAULT);
1257         }
1258
1259         if (unlikely(dlm_req->lock_desc.l_req_mode <= LCK_MINMODE ||
1260                      dlm_req->lock_desc.l_req_mode >= LCK_MAXMODE ||
1261                      dlm_req->lock_desc.l_req_mode &
1262                      (dlm_req->lock_desc.l_req_mode-1))) {
1263                 DEBUG_REQ(D_ERROR, req, "invalid lock request mode %d",
1264                           dlm_req->lock_desc.l_req_mode);
1265                 GOTO(out, rc = -EFAULT);
1266         }
1267
1268         if (unlikely((flags & LDLM_FL_REPLAY) ||
1269                      (lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT))) {
1270                 /* Find an existing lock in the per-export lock hash */
1271                 /*
1272                  * In the function below, .hs_keycmp resolves to
1273                  * ldlm_export_lock_keycmp()
1274                  */
1275                 /* coverity[overrun-buffer-val] */
1276                 lock = cfs_hash_lookup(req->rq_export->exp_lock_hash,
1277                                        (void *)&dlm_req->lock_handle[0]);
1278                 if (lock != NULL) {
1279                         DEBUG_REQ(D_DLMTRACE, req,
1280                                   "found existing lock cookie %#llx",
1281                                   lock->l_handle.h_cookie);
1282                         flags |= LDLM_FL_RESENT;
1283                         GOTO(existing_lock, rc = 0);
1284                 }
1285         } else {
1286                 if (ldlm_reclaim_full()) {
1287                         DEBUG_REQ(D_DLMTRACE, req,
1288                                   "Too many granted locks, reject current enqueue request and let the client retry later.\n");
1289                         GOTO(out, rc = -EINPROGRESS);
1290                 }
1291         }
1292
1293         /* The lock's callback data might be set in the policy function */
1294         lock = ldlm_lock_create(ns, &dlm_req->lock_desc.l_resource.lr_name,
1295                                 dlm_req->lock_desc.l_resource.lr_type,
1296                                 dlm_req->lock_desc.l_req_mode,
1297                                 cbs, NULL, 0, LVB_T_NONE);
1298         if (IS_ERR(lock)) {
1299                 rc = PTR_ERR(lock);
1300                 lock = NULL;
1301                 GOTO(out, rc);
1302         }
1303
1304         lock->l_remote_handle = dlm_req->lock_handle[0];
1305         LDLM_DEBUG(lock, "server-side enqueue handler, new lock created");
1306
1307         /*
1308          * Initialize resource lvb but not for a lock being replayed since
1309          * Client already got lvb sent in this case.
1310          * This must occur early since some policy methods assume resource
1311          * lvb is available (lr_lvb_data != NULL).
1312          */
1313         res = lock->l_resource;
1314         if (!(flags & LDLM_FL_REPLAY)) {
1315                 /* non-replayed lock, delayed lvb init may need to be done */
1316                 rc = ldlm_lvbo_init(res);
1317                 if (rc < 0) {
1318                         LDLM_DEBUG(lock, "delayed lvb init failed (rc %d)", rc);
1319                         GOTO(out, rc);
1320                 }
1321         }
1322
1323         OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_ENQUEUE_BLOCKED, obd_timeout * 2);
1324         /*
1325          * Don't enqueue a lock onto the export if it is been disonnected
1326          * due to eviction (b=3822) or server umount (b=24324).
1327          * Cancel it now instead.
1328          */
1329         if (req->rq_export->exp_disconnected) {
1330                 LDLM_ERROR(lock, "lock on disconnected export %p",
1331                            req->rq_export);
1332                 GOTO(out, rc = -ENOTCONN);
1333         }
1334
1335         lock->l_export = class_export_lock_get(req->rq_export, lock);
1336         if (lock->l_export->exp_lock_hash)
1337                 cfs_hash_add(lock->l_export->exp_lock_hash,
1338                              &lock->l_remote_handle,
1339                              &lock->l_exp_hash);
1340
1341         /*
1342          * Inherit the enqueue flags before the operation, because we do not
1343          * keep the res lock on return and next operations (BL AST) may proceed
1344          * without them.
1345          */
1346         lock->l_flags |= ldlm_flags_from_wire(dlm_req->lock_flags &
1347                                               LDLM_FL_INHERIT_MASK);
1348
1349         ldlm_convert_policy_to_local(req->rq_export,
1350                                      dlm_req->lock_desc.l_resource.lr_type,
1351                                      &dlm_req->lock_desc.l_policy_data,
1352                                      &lock->l_policy_data);
1353         if (dlm_req->lock_desc.l_resource.lr_type == LDLM_EXTENT)
1354                 lock->l_req_extent = lock->l_policy_data.l_extent;
1355
1356 existing_lock:
1357         if (flags & LDLM_FL_HAS_INTENT) {
1358                 /*
1359                  * In this case, the reply buffer is allocated deep in
1360                  * local_lock_enqueue by the policy function.
1361                  */
1362                 cookie = req;
1363         } else {
1364                 /*
1365                  * based on the assumption that lvb size never changes during
1366                  * resource life time otherwise it need resource->lr_lock's
1367                  * protection
1368                  */
1369                 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB,
1370                                      RCL_SERVER, ldlm_lvbo_size(lock));
1371
1372                 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_ENQUEUE_EXTENT_ERR))
1373                         GOTO(out, rc = -ENOMEM);
1374
1375                 rc = req_capsule_server_pack(&req->rq_pill);
1376                 if (rc)
1377                         GOTO(out, rc);
1378         }
1379
1380         err = ldlm_lock_enqueue(env, ns, &lock, cookie, &flags);
1381         if (err) {
1382                 if ((int)err < 0)
1383                         rc = (int)err;
1384                 GOTO(out, err);
1385         }
1386
1387         dlm_rep = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
1388
1389         ldlm_lock2desc(lock, &dlm_rep->lock_desc);
1390         ldlm_lock2handle(lock, &dlm_rep->lock_handle);
1391
1392         if (lock && lock->l_resource->lr_type == LDLM_EXTENT)
1393                 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_BL_EVICT, 6);
1394
1395         /*
1396          * We never send a blocking AST until the lock is granted, but
1397          * we can tell it right now
1398          */
1399         lock_res_and_lock(lock);
1400
1401         /*
1402          * Now take into account flags to be inherited from original lock
1403          * request both in reply to client and in our own lock flags.
1404          */
1405         dlm_rep->lock_flags = ldlm_flags_to_wire(flags);
1406         lock->l_flags |= flags & LDLM_FL_INHERIT_MASK;
1407
1408         /*
1409          * Don't move a pending lock onto the export if it has already been
1410          * disconnected due to eviction (b=5683) or server umount (b=24324).
1411          * Cancel it now instead.
1412          */
1413         if (unlikely(req->rq_export->exp_disconnected ||
1414                      OBD_FAIL_CHECK(OBD_FAIL_LDLM_ENQUEUE_OLD_EXPORT))) {
1415                 LDLM_ERROR(lock, "lock on destroyed export %p", req->rq_export);
1416                 rc = -ENOTCONN;
1417         } else if (ldlm_is_ast_sent(lock)) {
1418                 /* fill lock desc for possible lock convert */
1419                 if (lock->l_blocking_lock &&
1420                     lock->l_resource->lr_type == LDLM_IBITS) {
1421                         struct ldlm_lock *bl_lock = lock->l_blocking_lock;
1422                         struct ldlm_lock_desc *rep_desc = &dlm_rep->lock_desc;
1423
1424                         LDLM_DEBUG(lock,
1425                                    "save blocking bits %llx in granted lock",
1426                                    bl_lock->l_policy_data.l_inodebits.bits);
1427                         /*
1428                          * If lock is blocked then save blocking ibits
1429                          * in returned lock policy for the possible lock
1430                          * convert on a client.
1431                          */
1432                         rep_desc->l_policy_data.l_inodebits.cancel_bits =
1433                                 bl_lock->l_policy_data.l_inodebits.bits;
1434                 }
1435                 dlm_rep->lock_flags |= ldlm_flags_to_wire(LDLM_FL_AST_SENT);
1436                 if (ldlm_is_granted(lock)) {
1437                         /*
1438                          * Only cancel lock if it was granted, because it would
1439                          * be destroyed immediately and would never be granted
1440                          * in the future, causing timeouts on client.  Not
1441                          * granted lock will be cancelled immediately after
1442                          * sending completion AST.
1443                          */
1444                         if (ldlm_is_cancel_on_block(lock)) {
1445                                 unlock_res_and_lock(lock);
1446                                 ldlm_lock_cancel(lock);
1447                                 lock_res_and_lock(lock);
1448                         } else {
1449                                 ldlm_add_waiting_lock(lock,
1450                                                       ldlm_bl_timeout(lock));
1451                         }
1452                 }
1453         }
1454         unlock_res_and_lock(lock);
1455
1456         EXIT;
1457 out:
1458         req->rq_status = rc ?: err; /* return either error - b=11190 */
1459         if (!req->rq_packed_final) {
1460                 err = lustre_pack_reply(req, 1, NULL, NULL);
1461                 if (rc == 0)
1462                         rc = err;
1463         }
1464
1465         /*
1466          * The LOCK_CHANGED code in ldlm_lock_enqueue depends on this
1467          * ldlm_reprocess_all.  If this moves, revisit that code. -phil
1468          */
1469         if (lock != NULL) {
1470                 LDLM_DEBUG(lock,
1471                            "server-side enqueue handler, sending reply (err=%d, rc=%d)",
1472                            err, rc);
1473
1474                 if (rc == 0 &&
1475                     req_capsule_has_field(&req->rq_pill, &RMF_DLM_LVB,
1476                                           RCL_SERVER) &&
1477                     ldlm_lvbo_size(lock) > 0) {
1478                         void *buf;
1479                         int buflen;
1480
1481 retry:
1482                         buf = req_capsule_server_get(&req->rq_pill,
1483                                                      &RMF_DLM_LVB);
1484                         LASSERTF(buf != NULL, "req %p, lock %p\n", req, lock);
1485                         buflen = req_capsule_get_size(&req->rq_pill,
1486                                         &RMF_DLM_LVB, RCL_SERVER);
1487                         /*
1488                          * non-replayed lock, delayed lvb init may
1489                          * need to be occur now
1490                          */
1491                         if ((buflen > 0) && !(flags & LDLM_FL_REPLAY)) {
1492                                 int rc2;
1493
1494                                 rc2 = ldlm_lvbo_fill(lock, buf, &buflen);
1495                                 if (rc2 >= 0) {
1496                                         req_capsule_shrink(&req->rq_pill,
1497                                                            &RMF_DLM_LVB,
1498                                                            rc2, RCL_SERVER);
1499                                 } else if (rc2 == -ERANGE) {
1500                                         rc2 = req_capsule_server_grow(
1501                                                         &req->rq_pill,
1502                                                         &RMF_DLM_LVB, buflen);
1503                                         if (!rc2) {
1504                                                 goto retry;
1505                                         } else {
1506                                                 /*
1507                                                  * if we can't grow the buffer,
1508                                                  * it's ok to return empty lvb
1509                                                  * to client.
1510                                                  */
1511                                                 req_capsule_shrink(
1512                                                         &req->rq_pill,
1513                                                         &RMF_DLM_LVB, 0,
1514                                                         RCL_SERVER);
1515                                         }
1516                                 } else {
1517                                         rc = rc2;
1518                                 }
1519                         } else if (flags & LDLM_FL_REPLAY) {
1520                                 /* no LVB resend upon replay */
1521                                 if (buflen > 0)
1522                                         req_capsule_shrink(&req->rq_pill,
1523                                                            &RMF_DLM_LVB,
1524                                                            0, RCL_SERVER);
1525                                 else
1526                                         rc = buflen;
1527                         } else {
1528                                 rc = buflen;
1529                         }
1530                 }
1531
1532                 if (rc != 0 && !(flags & LDLM_FL_RESENT)) {
1533                         if (lock->l_export) {
1534                                 ldlm_lock_cancel(lock);
1535                         } else {
1536                                 lock_res_and_lock(lock);
1537                                 ldlm_resource_unlink_lock(lock);
1538                                 ldlm_lock_destroy_nolock(lock);
1539                                 unlock_res_and_lock(lock);
1540
1541                         }
1542                 }
1543
1544                 if (!err && !ldlm_is_cbpending(lock) &&
1545                     dlm_req->lock_desc.l_resource.lr_type != LDLM_FLOCK)
1546                         ldlm_reprocess_all(lock->l_resource, lock);
1547
1548                 LDLM_LOCK_RELEASE(lock);
1549         }
1550
1551         LDLM_DEBUG_NOLOCK("server-side enqueue handler END (lock %p, rc %d)",
1552                           lock, rc);
1553
1554         return rc;
1555 }
1556
1557 /*
1558  * Clear the blocking lock, the race is possible between ldlm_handle_convert0()
1559  * and ldlm_work_bl_ast_lock(), so this is done under lock with check for NULL.
1560  */
1561 void ldlm_clear_blocking_lock(struct ldlm_lock *lock)
1562 {
1563         if (lock->l_blocking_lock) {
1564                 LDLM_LOCK_RELEASE(lock->l_blocking_lock);
1565                 lock->l_blocking_lock = NULL;
1566         }
1567 }
1568
1569 /* A lock can be converted to new ibits or mode and should be considered
1570  * as new lock. Clear all states related to a previous blocking AST
1571  * processing so new conflicts will cause new blocking ASTs.
1572  *
1573  * This is used during lock convert below and lock downgrade to COS mode in
1574  * ldlm_lock_mode_downgrade().
1575  */
1576 void ldlm_clear_blocking_data(struct ldlm_lock *lock)
1577 {
1578         ldlm_clear_ast_sent(lock);
1579         lock->l_bl_ast_run = 0;
1580         ldlm_clear_blocking_lock(lock);
1581 }
1582
1583 /**
1584  * Main LDLM entry point for server code to process lock conversion requests.
1585  */
1586 int ldlm_handle_convert0(struct ptlrpc_request *req,
1587                          const struct ldlm_request *dlm_req)
1588 {
1589         struct obd_export *exp = req->rq_export;
1590         struct ldlm_reply *dlm_rep;
1591         struct ldlm_lock *lock;
1592         int rc;
1593
1594         ENTRY;
1595
1596         if (exp && exp->exp_nid_stats && exp->exp_nid_stats->nid_ldlm_stats)
1597                 lprocfs_counter_incr(exp->exp_nid_stats->nid_ldlm_stats,
1598                                      LDLM_CONVERT - LDLM_FIRST_OPC);
1599
1600         rc = req_capsule_server_pack(&req->rq_pill);
1601         if (rc)
1602                 RETURN(rc);
1603
1604         dlm_rep = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
1605         dlm_rep->lock_flags = dlm_req->lock_flags;
1606
1607         lock = ldlm_handle2lock(&dlm_req->lock_handle[0]);
1608         if (lock) {
1609                 __u64 bits;
1610                 __u64 new;
1611
1612                 bits = lock->l_policy_data.l_inodebits.bits;
1613                 new = dlm_req->lock_desc.l_policy_data.l_inodebits.bits;
1614                 LDLM_DEBUG(lock, "server-side convert handler START");
1615
1616                 if (ldlm_is_cancel(lock)) {
1617                         LDLM_ERROR(lock, "convert on canceled lock!");
1618                         rc = ELDLM_NO_LOCK_DATA;
1619                 } else if (dlm_req->lock_desc.l_req_mode !=
1620                            lock->l_granted_mode) {
1621                         LDLM_ERROR(lock, "lock mode differs!");
1622                         rc = ELDLM_NO_LOCK_DATA;
1623                 } else if (bits == new) {
1624                         /*
1625                          * This can be valid situation if CONVERT RPCs are
1626                          * re-ordered. Just finish silently
1627                          */
1628                         LDLM_DEBUG(lock, "lock is converted already!");
1629                         rc = ELDLM_OK;
1630                 } else {
1631                         lock_res_and_lock(lock);
1632                         if (ldlm_is_waited(lock))
1633                                 ldlm_del_waiting_lock(lock);
1634
1635                         ldlm_clear_cbpending(lock);
1636                         lock->l_policy_data.l_inodebits.cancel_bits = 0;
1637                         ldlm_inodebits_drop(lock, bits & ~new);
1638
1639                         ldlm_clear_blocking_data(lock);
1640                         unlock_res_and_lock(lock);
1641
1642                         ldlm_reprocess_all(lock->l_resource, NULL);
1643                         rc = ELDLM_OK;
1644                 }
1645
1646                 if (rc == ELDLM_OK) {
1647                         dlm_rep->lock_handle = lock->l_remote_handle;
1648                         ldlm_ibits_policy_local_to_wire(&lock->l_policy_data,
1649                                         &dlm_rep->lock_desc.l_policy_data);
1650                 }
1651
1652                 LDLM_DEBUG(lock, "server-side convert handler END, rc = %d",
1653                            rc);
1654                 LDLM_LOCK_PUT(lock);
1655         } else {
1656                 rc = ELDLM_NO_LOCK_DATA;
1657                 LDLM_DEBUG_NOLOCK("server-side convert handler END, rc = %d",
1658                                   rc);
1659         }
1660
1661         req->rq_status = rc;
1662
1663         RETURN(0);
1664 }
1665
1666 /**
1667  * Cancel all the locks whose handles are packed into ldlm_request
1668  *
1669  * Called by server code expecting such combined cancel activity
1670  * requests.
1671  */
1672 int ldlm_request_cancel(struct ptlrpc_request *req,
1673                         const struct ldlm_request *dlm_req,
1674                         int first, enum lustre_at_flags flags)
1675 {
1676         struct ldlm_resource *res, *pres = NULL;
1677         struct ldlm_lock *lock;
1678         int i, count, done = 0;
1679
1680         ENTRY;
1681
1682         count = dlm_req->lock_count ? dlm_req->lock_count : 1;
1683         if (first >= count)
1684                 RETURN(0);
1685
1686         if (count == 1 && dlm_req->lock_handle[0].cookie == 0)
1687                 RETURN(0);
1688
1689         /*
1690          * There is no lock on the server at the replay time,
1691          * skip lock cancelling to make replay tests to pass.
1692          */
1693         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
1694                 RETURN(0);
1695
1696         LDLM_DEBUG_NOLOCK("server-side cancel handler START: %d locks, starting at %d",
1697                           count, first);
1698
1699         for (i = first; i < count; i++) {
1700                 lock = ldlm_handle2lock(&dlm_req->lock_handle[i]);
1701                 if (!lock) {
1702                         LDLM_DEBUG_NOLOCK("server-side cancel handler stale lock (cookie %llu)",
1703                                           dlm_req->lock_handle[i].cookie);
1704                         continue;
1705                 }
1706
1707                 res = lock->l_resource;
1708                 done++;
1709
1710                 /*
1711                  * This code is an optimization to only attempt lock
1712                  * granting on the resource (that could be CPU-expensive)
1713                  * after we are done cancelling lock in that resource.
1714                  */
1715                 if (res != pres) {
1716                         if (pres != NULL) {
1717                                 ldlm_reprocess_all(pres, NULL);
1718                                 LDLM_RESOURCE_DELREF(pres);
1719                                 ldlm_resource_putref(pres);
1720                         }
1721                         if (res != NULL) {
1722                                 ldlm_resource_getref(res);
1723                                 LDLM_RESOURCE_ADDREF(res);
1724
1725                                 if (!ldlm_is_discard_data(lock))
1726                                         ldlm_lvbo_update(res, lock,
1727                                                          NULL, 1);
1728                         }
1729                         pres = res;
1730                 }
1731
1732                 if ((flags & LATF_STATS) && ldlm_is_ast_sent(lock) &&
1733                     lock->l_blast_sent != 0) {
1734                         time64_t delay = ktime_get_real_seconds() -
1735                                          lock->l_blast_sent;
1736                         LDLM_DEBUG(lock,
1737                                    "server cancels blocked lock after %llds",
1738                                    (s64)delay);
1739                         at_measured(&lock->l_export->exp_bl_lock_at, delay);
1740                 }
1741                 ldlm_lock_cancel(lock);
1742                 LDLM_LOCK_PUT(lock);
1743         }
1744         if (pres != NULL) {
1745                 ldlm_reprocess_all(pres, NULL);
1746                 LDLM_RESOURCE_DELREF(pres);
1747                 ldlm_resource_putref(pres);
1748         }
1749         LDLM_DEBUG_NOLOCK("server-side cancel handler END");
1750         RETURN(done);
1751 }
1752 EXPORT_SYMBOL(ldlm_request_cancel);
1753
1754 /**
1755  * Main LDLM entry point for server code to cancel locks.
1756  *
1757  * Typically gets called from service handler on LDLM_CANCEL opc.
1758  */
1759 int ldlm_handle_cancel(struct ptlrpc_request *req)
1760 {
1761         struct ldlm_request *dlm_req;
1762         int rc;
1763
1764         ENTRY;
1765
1766         dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1767         if (dlm_req == NULL) {
1768                 CDEBUG(D_INFO, "bad request buffer for cancel\n");
1769                 RETURN(-EFAULT);
1770         }
1771
1772         if (req->rq_export && req->rq_export->exp_nid_stats &&
1773             req->rq_export->exp_nid_stats->nid_ldlm_stats)
1774                 lprocfs_counter_incr(req->rq_export->exp_nid_stats->nid_ldlm_stats,
1775                                      LDLM_CANCEL - LDLM_FIRST_OPC);
1776
1777         rc = req_capsule_server_pack(&req->rq_pill);
1778         if (rc)
1779                 RETURN(rc);
1780
1781         if (!ldlm_request_cancel(req, dlm_req, 0, LATF_STATS))
1782                 req->rq_status = LUSTRE_ESTALE;
1783
1784         RETURN(ptlrpc_reply(req));
1785 }
1786 #endif /* HAVE_SERVER_SUPPORT */
1787
1788 /**
1789  * Callback handler for receiving incoming blocking ASTs.
1790  *
1791  * This can only happen on client side.
1792  */
1793 void ldlm_handle_bl_callback(struct ldlm_namespace *ns,
1794                              struct ldlm_lock_desc *ld, struct ldlm_lock *lock)
1795 {
1796         int do_ast;
1797
1798         ENTRY;
1799
1800         LDLM_DEBUG(lock, "client blocking AST callback handler");
1801
1802         lock_res_and_lock(lock);
1803
1804         /* set bits to cancel for this lock for possible lock convert */
1805         if (ns_is_client(ns) && (lock->l_resource->lr_type == LDLM_IBITS)) {
1806                 /*
1807                  * Lock description contains policy of blocking lock,
1808                  * and its cancel_bits is used to pass conflicting bits.
1809                  * NOTE: ld can be NULL or can be not NULL but zeroed if
1810                  * passed from ldlm_bl_thread_blwi(), check below used bits
1811                  * in ld to make sure it is valid description.
1812                  *
1813                  * If server may replace lock resource keeping the same cookie,
1814                  * never use cancel bits from different resource, full cancel
1815                  * is to be used.
1816                  */
1817                 if (ld && ld->l_policy_data.l_inodebits.bits &&
1818                     ldlm_res_eq(&ld->l_resource.lr_name,
1819                                 &lock->l_resource->lr_name))
1820                         lock->l_policy_data.l_inodebits.cancel_bits =
1821                                 ld->l_policy_data.l_inodebits.cancel_bits;
1822                 /*
1823                  * if there is no valid ld and lock is cbpending already
1824                  * then cancel_bits should be kept, otherwise it is zeroed.
1825                  */
1826                 else if (!ldlm_is_cbpending(lock))
1827                         lock->l_policy_data.l_inodebits.cancel_bits = 0;
1828         }
1829         ldlm_set_cbpending(lock);
1830
1831         do_ast = (!lock->l_readers && !lock->l_writers);
1832         unlock_res_and_lock(lock);
1833
1834         if (do_ast) {
1835                 CDEBUG(D_DLMTRACE,
1836                        "Lock %p already unused, calling callback (%p)\n",
1837                        lock, lock->l_blocking_ast);
1838                 if (lock->l_blocking_ast != NULL)
1839                         lock->l_blocking_ast(lock, ld, lock->l_ast_data,
1840                                              LDLM_CB_BLOCKING);
1841         } else {
1842                 CDEBUG(D_DLMTRACE,
1843                        "Lock %p is referenced, will be cancelled later\n",
1844                        lock);
1845         }
1846
1847         LDLM_DEBUG(lock, "client blocking callback handler END");
1848         LDLM_LOCK_RELEASE(lock);
1849         EXIT;
1850 }
1851
1852 /**
1853  * Callback handler for receiving incoming completion ASTs.
1854  *
1855  * This only can happen on client side.
1856  */
1857 static void ldlm_handle_cp_callback(struct ptlrpc_request *req,
1858                                     struct ldlm_namespace *ns,
1859                                     struct ldlm_request *dlm_req,
1860                                     struct ldlm_lock *lock)
1861 {
1862         struct list_head ast_list;
1863         int lvb_len;
1864         int rc = 0;
1865
1866         ENTRY;
1867
1868         LDLM_DEBUG(lock, "client completion callback handler START");
1869
1870         INIT_LIST_HEAD(&ast_list);
1871         if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_BL_CB_RACE)) {
1872                 long to = cfs_time_seconds(1);
1873
1874                 while (to > 0) {
1875                         set_current_state(TASK_INTERRUPTIBLE);
1876                         schedule_timeout(to);
1877                         if (ldlm_is_granted(lock) ||
1878                             ldlm_is_destroyed(lock))
1879                                 break;
1880                 }
1881         }
1882
1883         lvb_len = req_capsule_get_size(&req->rq_pill, &RMF_DLM_LVB, RCL_CLIENT);
1884         if (lvb_len < 0) {
1885                 LDLM_ERROR(lock, "Fail to get lvb_len, rc = %d", lvb_len);
1886                 GOTO(out, rc = lvb_len);
1887         } else if (lvb_len > 0) {
1888                 if (lock->l_lvb_len > 0) {
1889                         /* for extent lock, lvb contains ost_lvb{}. */
1890                         LASSERT(lock->l_lvb_data != NULL);
1891
1892                         if (unlikely(lock->l_lvb_len < lvb_len)) {
1893                                 LDLM_ERROR(lock,
1894                                            "Replied LVB is larger than expectation, expected = %d, replied = %d",
1895                                            lock->l_lvb_len, lvb_len);
1896                                 GOTO(out, rc = -EINVAL);
1897                         }
1898                 }
1899         }
1900
1901         lock_res_and_lock(lock);
1902
1903         if (!ldlm_res_eq(&dlm_req->lock_desc.l_resource.lr_name,
1904                          &lock->l_resource->lr_name)) {
1905                 ldlm_resource_unlink_lock(lock);
1906                 unlock_res_and_lock(lock);
1907                 rc = ldlm_lock_change_resource(ns, lock,
1908                                 &dlm_req->lock_desc.l_resource.lr_name);
1909                 if (rc < 0) {
1910                         LDLM_ERROR(lock, "Failed to allocate resource");
1911                         GOTO(out, rc);
1912                 }
1913                 LDLM_DEBUG(lock, "completion AST, new resource");
1914                 lock_res_and_lock(lock);
1915         }
1916
1917         if (ldlm_is_destroyed(lock) ||
1918             ldlm_is_granted(lock)) {
1919                 /* b=11300: the lock has already been granted */
1920                 unlock_res_and_lock(lock);
1921                 LDLM_DEBUG(lock, "Double grant race happened");
1922                 GOTO(out, rc = 0);
1923         }
1924
1925         /*
1926          * If we receive the completion AST before the actual enqueue returned,
1927          * then we might need to switch lock modes, resources, or extents.
1928          */
1929         if (dlm_req->lock_desc.l_granted_mode != lock->l_req_mode) {
1930                 lock->l_req_mode = dlm_req->lock_desc.l_granted_mode;
1931                 LDLM_DEBUG(lock, "completion AST, new lock mode");
1932         }
1933
1934         if (lock->l_resource->lr_type != LDLM_PLAIN) {
1935                 ldlm_convert_policy_to_local(req->rq_export,
1936                                           dlm_req->lock_desc.l_resource.lr_type,
1937                                           &dlm_req->lock_desc.l_policy_data,
1938                                           &lock->l_policy_data);
1939                 LDLM_DEBUG(lock, "completion AST, new policy data");
1940         }
1941
1942         ldlm_resource_unlink_lock(lock);
1943
1944         if (dlm_req->lock_flags & LDLM_FL_AST_SENT) {
1945                 /*
1946                  * BL_AST locks are not needed in LRU.
1947                  * Let ldlm_cancel_lru() be fast.
1948                  */
1949                 ldlm_lock_remove_from_lru(lock);
1950                 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_BL_AST;
1951                 LDLM_DEBUG(lock, "completion AST includes blocking AST");
1952         }
1953
1954         if (lock->l_lvb_len > 0) {
1955                 rc = ldlm_fill_lvb(lock, &req->rq_pill, RCL_CLIENT,
1956                                    lock->l_lvb_data, lvb_len);
1957                 if (rc < 0) {
1958                         unlock_res_and_lock(lock);
1959                         GOTO(out, rc);
1960                 }
1961         }
1962
1963         ldlm_grant_lock(lock, &ast_list);
1964         unlock_res_and_lock(lock);
1965
1966         LDLM_DEBUG(lock, "callback handler finished, about to run_ast_work");
1967
1968         /*
1969          * Let Enqueue to call osc_lock_upcall() and initialize
1970          * l_ast_data
1971          */
1972         OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_CP_ENQ_RACE, 2);
1973
1974         ldlm_run_ast_work(ns, &ast_list, LDLM_WORK_CP_AST);
1975
1976         LDLM_DEBUG_NOLOCK("client completion callback handler END (lock %p)",
1977                           lock);
1978         GOTO(out, rc);
1979
1980 out:
1981         if (rc < 0) {
1982                 lock_res_and_lock(lock);
1983                 ldlm_set_failed(lock);
1984                 unlock_res_and_lock(lock);
1985                 wake_up(&lock->l_waitq);
1986         }
1987         LDLM_LOCK_RELEASE(lock);
1988 }
1989
1990 /**
1991  * Callback handler for receiving incoming glimpse ASTs.
1992  *
1993  * This only can happen on client side.  After handling the glimpse AST
1994  * we also consider dropping the lock here if it is unused locally for a
1995  * long time.
1996  */
1997 static void ldlm_handle_gl_callback(struct ptlrpc_request *req,
1998                                     struct ldlm_namespace *ns,
1999                                     struct ldlm_request *dlm_req,
2000                                     struct ldlm_lock *lock)
2001 {
2002         int rc = -ENOSYS;
2003
2004         ENTRY;
2005
2006         LDLM_DEBUG(lock, "client glimpse AST callback handler");
2007
2008         if (lock->l_glimpse_ast != NULL)
2009                 rc = lock->l_glimpse_ast(lock, req);
2010
2011         if (req->rq_repmsg != NULL) {
2012                 ptlrpc_reply(req);
2013         } else {
2014                 req->rq_status = rc;
2015                 ptlrpc_error(req);
2016         }
2017
2018         lock_res_and_lock(lock);
2019         if (lock->l_granted_mode == LCK_PW &&
2020             !lock->l_readers && !lock->l_writers &&
2021             ktime_after(ktime_get(),
2022                         ktime_add(lock->l_last_used,
2023                                   ktime_set(ns->ns_dirty_age_limit, 0)))) {
2024                 unlock_res_and_lock(lock);
2025                 if (ldlm_bl_to_thread_lock(ns, NULL, lock))
2026                         ldlm_handle_bl_callback(ns, NULL, lock);
2027
2028                 EXIT;
2029                 return;
2030         }
2031         unlock_res_and_lock(lock);
2032         LDLM_LOCK_RELEASE(lock);
2033         EXIT;
2034 }
2035
2036 static int ldlm_callback_reply(struct ptlrpc_request *req, int rc)
2037 {
2038         if (req->rq_no_reply)
2039                 return 0;
2040
2041         req->rq_status = rc;
2042         if (!req->rq_packed_final) {
2043                 rc = lustre_pack_reply(req, 1, NULL, NULL);
2044                 if (rc)
2045                         return rc;
2046         }
2047         return ptlrpc_reply(req);
2048 }
2049
2050 static int __ldlm_bl_to_thread(struct ldlm_bl_work_item *blwi,
2051                                enum ldlm_cancel_flags cancel_flags)
2052 {
2053         struct ldlm_bl_pool *blp = ldlm_state->ldlm_bl_pool;
2054
2055         ENTRY;
2056
2057         spin_lock(&blp->blp_lock);
2058         if (blwi->blwi_lock &&
2059             ldlm_is_discard_data(blwi->blwi_lock)) {
2060                 /* add LDLM_FL_DISCARD_DATA requests to the priority list */
2061                 list_add_tail(&blwi->blwi_entry, &blp->blp_prio_list);
2062         } else {
2063                 /* other blocking callbacks are added to the regular list */
2064                 list_add_tail(&blwi->blwi_entry, &blp->blp_list);
2065         }
2066         spin_unlock(&blp->blp_lock);
2067
2068         wake_up(&blp->blp_waitq);
2069
2070         /*
2071          * can not check blwi->blwi_flags as blwi could be already freed in
2072          * LCF_ASYNC mode
2073          */
2074         if (!(cancel_flags & LCF_ASYNC))
2075                 wait_for_completion(&blwi->blwi_comp);
2076
2077         RETURN(0);
2078 }
2079
2080 static inline void init_blwi(struct ldlm_bl_work_item *blwi,
2081                              struct ldlm_namespace *ns,
2082                              struct ldlm_lock_desc *ld,
2083                              struct list_head *cancels, int count,
2084                              struct ldlm_lock *lock,
2085                              enum ldlm_cancel_flags cancel_flags)
2086 {
2087         init_completion(&blwi->blwi_comp);
2088         INIT_LIST_HEAD(&blwi->blwi_head);
2089
2090         if (memory_pressure_get())
2091                 blwi->blwi_mem_pressure = 1;
2092
2093         blwi->blwi_ns = ns;
2094         blwi->blwi_flags = cancel_flags;
2095         if (ld != NULL)
2096                 blwi->blwi_ld = *ld;
2097         if (count) {
2098                 list_add(&blwi->blwi_head, cancels);
2099                 list_del_init(cancels);
2100                 blwi->blwi_count = count;
2101         } else {
2102                 blwi->blwi_lock = lock;
2103         }
2104 }
2105
2106 /**
2107  * Queues a list of locks \a cancels containing \a count locks
2108  * for later processing by a blocking thread.  If \a count is zero,
2109  * then the lock referenced as \a lock is queued instead.
2110  *
2111  * The blocking thread would then call ->l_blocking_ast callback in the lock.
2112  * If list addition fails an error is returned and caller is supposed to
2113  * call ->l_blocking_ast itself.
2114  */
2115 static int ldlm_bl_to_thread(struct ldlm_namespace *ns,
2116                              struct ldlm_lock_desc *ld,
2117                              struct ldlm_lock *lock,
2118                              struct list_head *cancels, int count,
2119                              enum ldlm_cancel_flags cancel_flags)
2120 {
2121         ENTRY;
2122
2123         if (cancels && count == 0)
2124                 RETURN(0);
2125
2126         if (cancel_flags & LCF_ASYNC) {
2127                 struct ldlm_bl_work_item *blwi;
2128
2129                 OBD_ALLOC(blwi, sizeof(*blwi));
2130                 if (blwi == NULL)
2131                         RETURN(-ENOMEM);
2132                 init_blwi(blwi, ns, ld, cancels, count, lock, cancel_flags);
2133
2134                 RETURN(__ldlm_bl_to_thread(blwi, cancel_flags));
2135         } else {
2136                 /*
2137                  * if it is synchronous call do minimum mem alloc, as it could
2138                  * be triggered from kernel shrinker
2139                  */
2140                 struct ldlm_bl_work_item blwi;
2141
2142                 memset(&blwi, 0, sizeof(blwi));
2143                 init_blwi(&blwi, ns, ld, cancels, count, lock, cancel_flags);
2144                 RETURN(__ldlm_bl_to_thread(&blwi, cancel_flags));
2145         }
2146 }
2147
2148
2149 int ldlm_bl_to_thread_lock(struct ldlm_namespace *ns, struct ldlm_lock_desc *ld,
2150                            struct ldlm_lock *lock)
2151 {
2152         return ldlm_bl_to_thread(ns, ld, lock, NULL, 0, LCF_ASYNC);
2153 }
2154
2155 int ldlm_bl_to_thread_list(struct ldlm_namespace *ns, struct ldlm_lock_desc *ld,
2156                            struct list_head *cancels, int count,
2157                            enum ldlm_cancel_flags cancel_flags)
2158 {
2159         return ldlm_bl_to_thread(ns, ld, NULL, cancels, count, cancel_flags);
2160 }
2161
2162 int ldlm_bl_thread_wakeup(void)
2163 {
2164         wake_up(&ldlm_state->ldlm_bl_pool->blp_waitq);
2165         return 0;
2166 }
2167
2168 /* Setinfo coming from Server (eg MDT) to Client (eg MDC)! */
2169 static int ldlm_handle_setinfo(struct ptlrpc_request *req)
2170 {
2171         struct obd_device *obd = req->rq_export->exp_obd;
2172         char *key;
2173         void *val;
2174         int keylen, vallen;
2175         int rc = -ENOSYS;
2176
2177         ENTRY;
2178
2179         DEBUG_REQ(D_HSM, req, "%s: handle setinfo\n", obd->obd_name);
2180
2181         req_capsule_set(&req->rq_pill, &RQF_OBD_SET_INFO);
2182
2183         key = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_KEY);
2184         if (key == NULL) {
2185                 DEBUG_REQ(D_IOCTL, req, "no set_info key");
2186                 RETURN(-EFAULT);
2187         }
2188         keylen = req_capsule_get_size(&req->rq_pill, &RMF_SETINFO_KEY,
2189                                       RCL_CLIENT);
2190         val = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_VAL);
2191         if (val == NULL) {
2192                 DEBUG_REQ(D_IOCTL, req, "no set_info val");
2193                 RETURN(-EFAULT);
2194         }
2195         vallen = req_capsule_get_size(&req->rq_pill, &RMF_SETINFO_VAL,
2196                                       RCL_CLIENT);
2197
2198         /* We are responsible for swabbing contents of val */
2199
2200         if (KEY_IS(KEY_HSM_COPYTOOL_SEND))
2201                 /* Pass it on to mdc (the "export" in this case) */
2202                 rc = obd_set_info_async(req->rq_svc_thread->t_env,
2203                                         req->rq_export,
2204                                         sizeof(KEY_HSM_COPYTOOL_SEND),
2205                                         KEY_HSM_COPYTOOL_SEND,
2206                                         vallen, val, NULL);
2207         else
2208                 DEBUG_REQ(D_WARNING, req, "ignoring unknown key %s", key);
2209
2210         return rc;
2211 }
2212
2213 static inline void ldlm_callback_errmsg(struct ptlrpc_request *req,
2214                                         const char *msg, int rc,
2215                                         const struct lustre_handle *handle)
2216 {
2217         DEBUG_REQ((req->rq_no_reply || rc) ? D_WARNING : D_DLMTRACE, req,
2218                   "%s: [nid %s] [rc %d] [lock %#llx]",
2219                   msg, libcfs_id2str(req->rq_peer), rc,
2220                   handle ? handle->cookie : 0);
2221         if (req->rq_no_reply)
2222                 CWARN("No reply was sent, maybe cause b=21636.\n");
2223         else if (rc)
2224                 CWARN("Send reply failed, maybe cause b=21636.\n");
2225 }
2226
2227 /* TODO: handle requests in a similar way as MDT: see mdt_handle_common() */
2228 static int ldlm_callback_handler(struct ptlrpc_request *req)
2229 {
2230         struct ldlm_namespace *ns;
2231         struct ldlm_request *dlm_req;
2232         struct ldlm_lock *lock;
2233         int rc;
2234
2235         ENTRY;
2236
2237         /*
2238          * Requests arrive in sender's byte order.  The ptlrpc service
2239          * handler has already checked and, if necessary, byte-swapped the
2240          * incoming request message body, but I am responsible for the
2241          * message buffers.
2242          */
2243
2244         /* do nothing for sec context finalize */
2245         if (lustre_msg_get_opc(req->rq_reqmsg) == SEC_CTX_FINI)
2246                 RETURN(0);
2247
2248         req_capsule_init(&req->rq_pill, req, RCL_SERVER);
2249
2250         if (req->rq_export == NULL) {
2251                 rc = ldlm_callback_reply(req, -ENOTCONN);
2252                 ldlm_callback_errmsg(req, "Operate on unconnected server",
2253                                      rc, NULL);
2254                 RETURN(0);
2255         }
2256
2257         LASSERT(req->rq_export != NULL);
2258         LASSERT(req->rq_export->exp_obd != NULL);
2259
2260         switch (lustre_msg_get_opc(req->rq_reqmsg)) {
2261         case LDLM_BL_CALLBACK:
2262                 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_BL_CALLBACK_NET)) {
2263                         if (cfs_fail_err)
2264                                 ldlm_callback_reply(req, -(int)cfs_fail_err);
2265                         RETURN(0);
2266                 }
2267                 break;
2268         case LDLM_CP_CALLBACK:
2269                 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CP_CALLBACK_NET))
2270                         RETURN(0);
2271                 break;
2272         case LDLM_GL_CALLBACK:
2273                 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_GL_CALLBACK_NET))
2274                         RETURN(0);
2275                 break;
2276         case LDLM_SET_INFO:
2277                 rc = ldlm_handle_setinfo(req);
2278                 ldlm_callback_reply(req, rc);
2279                 RETURN(0);
2280         default:
2281                 CERROR("unknown opcode %u\n",
2282                        lustre_msg_get_opc(req->rq_reqmsg));
2283                 ldlm_callback_reply(req, -EPROTO);
2284                 RETURN(0);
2285         }
2286
2287         ns = req->rq_export->exp_obd->obd_namespace;
2288         LASSERT(ns != NULL);
2289
2290         req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK);
2291
2292         dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2293         if (dlm_req == NULL) {
2294                 rc = ldlm_callback_reply(req, -EPROTO);
2295                 ldlm_callback_errmsg(req, "Operate without parameter", rc,
2296                                      NULL);
2297                 RETURN(0);
2298         }
2299
2300         /*
2301          * Force a known safe race, send a cancel to the server for a lock
2302          * which the server has already started a blocking callback on.
2303          */
2304         if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_BL_CB_RACE) &&
2305             lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK) {
2306                 rc = ldlm_cli_cancel(&dlm_req->lock_handle[0], 0);
2307                 if (rc < 0)
2308                         CERROR("ldlm_cli_cancel: %d\n", rc);
2309         }
2310
2311         lock = ldlm_handle2lock_long(&dlm_req->lock_handle[0], 0);
2312         if (!lock) {
2313                 CDEBUG(D_DLMTRACE,
2314                        "callback on lock %#llx - lock disappeared\n",
2315                        dlm_req->lock_handle[0].cookie);
2316                 rc = ldlm_callback_reply(req, -EINVAL);
2317                 ldlm_callback_errmsg(req, "Operate with invalid parameter", rc,
2318                                      &dlm_req->lock_handle[0]);
2319                 RETURN(0);
2320         }
2321
2322         if (ldlm_is_fail_loc(lock) &&
2323             lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK)
2324                 OBD_RACE(OBD_FAIL_LDLM_CP_BL_RACE);
2325
2326         /* Copy hints/flags (e.g. LDLM_FL_DISCARD_DATA) from AST. */
2327         lock_res_and_lock(lock);
2328         lock->l_flags |= ldlm_flags_from_wire(dlm_req->lock_flags &
2329                                               LDLM_FL_AST_MASK);
2330         if (lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK) {
2331                 /*
2332                  * If somebody cancels lock and cache is already dropped,
2333                  * or lock is failed before cp_ast received on client,
2334                  * we can tell the server we have no lock. Otherwise, we
2335                  * should send cancel after dropping the cache.
2336                  */
2337                 if ((ldlm_is_canceling(lock) && ldlm_is_bl_done(lock)) ||
2338                      ldlm_is_failed(lock)) {
2339                         LDLM_DEBUG(lock,
2340                                    "callback on lock %llx - lock disappeared",
2341                                    dlm_req->lock_handle[0].cookie);
2342                         unlock_res_and_lock(lock);
2343                         LDLM_LOCK_RELEASE(lock);
2344                         rc = ldlm_callback_reply(req, -EINVAL);
2345                         ldlm_callback_errmsg(req, "Operate on stale lock", rc,
2346                                              &dlm_req->lock_handle[0]);
2347                         RETURN(0);
2348                 }
2349                 /*
2350                  * BL_AST locks are not needed in LRU.
2351                  * Let ldlm_cancel_lru() be fast.
2352                  */
2353                 ldlm_lock_remove_from_lru(lock);
2354                 ldlm_set_bl_ast(lock);
2355         }
2356         unlock_res_and_lock(lock);
2357
2358         /*
2359          * We want the ost thread to get this reply so that it can respond
2360          * to ost requests (write cache writeback) that might be triggered
2361          * in the callback.
2362          *
2363          * But we'd also like to be able to indicate in the reply that we're
2364          * cancelling right now, because it's unused, or have an intent result
2365          * in the reply, so we might have to push the responsibility for sending
2366          * the reply down into the AST handlers, alas.
2367          */
2368
2369         switch (lustre_msg_get_opc(req->rq_reqmsg)) {
2370         case LDLM_BL_CALLBACK:
2371                 CDEBUG(D_INODE, "blocking ast\n");
2372                 req_capsule_extend(&req->rq_pill, &RQF_LDLM_BL_CALLBACK);
2373                 if (!ldlm_is_cancel_on_block(lock)) {
2374                         rc = ldlm_callback_reply(req, 0);
2375                         if (req->rq_no_reply || rc)
2376                                 ldlm_callback_errmsg(req, "Normal process", rc,
2377                                                      &dlm_req->lock_handle[0]);
2378                 }
2379                 if (ldlm_bl_to_thread_lock(ns, &dlm_req->lock_desc, lock))
2380                         ldlm_handle_bl_callback(ns, &dlm_req->lock_desc, lock);
2381                 break;
2382         case LDLM_CP_CALLBACK:
2383                 CDEBUG(D_INODE, "completion ast\n");
2384                 req_capsule_extend(&req->rq_pill, &RQF_LDLM_CP_CALLBACK);
2385                 ldlm_callback_reply(req, 0);
2386                 ldlm_handle_cp_callback(req, ns, dlm_req, lock);
2387                 break;
2388         case LDLM_GL_CALLBACK:
2389                 CDEBUG(D_INODE, "glimpse ast\n");
2390                 req_capsule_extend(&req->rq_pill, &RQF_LDLM_GL_CALLBACK);
2391                 ldlm_handle_gl_callback(req, ns, dlm_req, lock);
2392                 break;
2393         default:
2394                 LBUG(); /* checked above */
2395         }
2396
2397         RETURN(0);
2398 }
2399
2400 #ifdef HAVE_SERVER_SUPPORT
2401 /**
2402  * Main handler for canceld thread.
2403  *
2404  * Separated into its own thread to avoid deadlocks.
2405  */
2406 static int ldlm_cancel_handler(struct ptlrpc_request *req)
2407 {
2408         int rc;
2409
2410         ENTRY;
2411
2412         /*
2413          * Requests arrive in sender's byte order.  The ptlrpc service
2414          * handler has already checked and, if necessary, byte-swapped the
2415          * incoming request message body, but I am responsible for the
2416          * message buffers.
2417          */
2418
2419         req_capsule_init(&req->rq_pill, req, RCL_SERVER);
2420
2421         if (req->rq_export == NULL) {
2422                 struct ldlm_request *dlm_req;
2423
2424                 CERROR("%s from %s arrived at %llu with bad export cookie %llu\n",
2425                        ll_opcode2str(lustre_msg_get_opc(req->rq_reqmsg)),
2426                        libcfs_nid2str(req->rq_peer.nid),
2427                        (unsigned long long)req->rq_arrival_time.tv_sec,
2428                        lustre_msg_get_handle(req->rq_reqmsg)->cookie);
2429
2430                 if (lustre_msg_get_opc(req->rq_reqmsg) == LDLM_CANCEL) {
2431                         req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK);
2432                         dlm_req = req_capsule_client_get(&req->rq_pill,
2433                                                          &RMF_DLM_REQ);
2434                         if (dlm_req != NULL)
2435                                 ldlm_lock_dump_handle(D_ERROR,
2436                                                       &dlm_req->lock_handle[0]);
2437                 }
2438                 ldlm_callback_reply(req, -ENOTCONN);
2439                 RETURN(0);
2440         }
2441
2442         switch (lustre_msg_get_opc(req->rq_reqmsg)) {
2443         /* XXX FIXME move this back to mds/handler.c, b=249 */
2444         case LDLM_CANCEL:
2445                 req_capsule_set(&req->rq_pill, &RQF_LDLM_CANCEL);
2446                 CDEBUG(D_INODE, "cancel\n");
2447                 if (CFS_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_NET) ||
2448                     CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_CANCEL_RESEND) ||
2449                     CFS_FAIL_CHECK(OBD_FAIL_LDLM_BL_EVICT))
2450                         RETURN(0);
2451                 rc = ldlm_handle_cancel(req);
2452                 break;
2453         case LDLM_CONVERT:
2454         {
2455                 struct ldlm_request *dlm_req;
2456
2457                 req_capsule_set(&req->rq_pill, &RQF_LDLM_CONVERT);
2458                 CDEBUG(D_INODE, "convert\n");
2459
2460                 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2461                 if (dlm_req == NULL) {
2462                         CDEBUG(D_INFO, "bad request buffer for cancel\n");
2463                         rc = ldlm_callback_reply(req, -EPROTO);
2464                 } else {
2465                         req->rq_status = ldlm_handle_convert0(req, dlm_req);
2466                         rc = ptlrpc_reply(req);
2467                 }
2468                 break;
2469         }
2470         default:
2471                 CERROR("invalid opcode %d\n",
2472                        lustre_msg_get_opc(req->rq_reqmsg));
2473                 req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK);
2474                 rc = ldlm_callback_reply(req, -EINVAL);
2475         }
2476
2477         RETURN(rc);
2478 }
2479
2480 static int ldlm_cancel_hpreq_lock_match(struct ptlrpc_request *req,
2481                                         struct ldlm_lock *lock)
2482 {
2483         struct ldlm_request *dlm_req;
2484         struct lustre_handle lockh;
2485         int rc = 0;
2486         int i;
2487
2488         ENTRY;
2489
2490         dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2491         if (dlm_req == NULL)
2492                 RETURN(0);
2493
2494         ldlm_lock2handle(lock, &lockh);
2495         for (i = 0; i < dlm_req->lock_count; i++) {
2496                 if (lustre_handle_equal(&dlm_req->lock_handle[i],
2497                                         &lockh)) {
2498                         DEBUG_REQ(D_RPCTRACE, req,
2499                                   "Prio raised by lock %#llx.", lockh.cookie);
2500                         rc = 1;
2501                         break;
2502                 }
2503         }
2504
2505         RETURN(rc);
2506 }
2507
2508 static int ldlm_cancel_hpreq_check(struct ptlrpc_request *req)
2509 {
2510         struct ldlm_request *dlm_req;
2511         int rc = 0;
2512         int i;
2513
2514         ENTRY;
2515
2516         /* no prolong in recovery */
2517         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
2518                 RETURN(0);
2519
2520         dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2521         if (dlm_req == NULL)
2522                 RETURN(-EFAULT);
2523
2524         for (i = 0; i < dlm_req->lock_count; i++) {
2525                 struct ldlm_lock *lock;
2526
2527                 lock = ldlm_handle2lock(&dlm_req->lock_handle[i]);
2528                 if (lock == NULL)
2529                         continue;
2530
2531                 rc = ldlm_is_ast_sent(lock) ? 1 : 0;
2532                 if (rc)
2533                         LDLM_DEBUG(lock, "hpreq cancel/convert lock");
2534                 LDLM_LOCK_PUT(lock);
2535
2536                 if (rc)
2537                         break;
2538         }
2539
2540         RETURN(rc);
2541 }
2542
2543 static struct ptlrpc_hpreq_ops ldlm_cancel_hpreq_ops = {
2544         .hpreq_lock_match = ldlm_cancel_hpreq_lock_match,
2545         .hpreq_check      = ldlm_cancel_hpreq_check,
2546         .hpreq_fini       = NULL,
2547 };
2548
2549 static int ldlm_hpreq_handler(struct ptlrpc_request *req)
2550 {
2551         ENTRY;
2552
2553         req_capsule_init(&req->rq_pill, req, RCL_SERVER);
2554
2555         if (req->rq_export == NULL)
2556                 RETURN(0);
2557
2558         if (LDLM_CANCEL == lustre_msg_get_opc(req->rq_reqmsg)) {
2559                 req_capsule_set(&req->rq_pill, &RQF_LDLM_CANCEL);
2560                 req->rq_ops = &ldlm_cancel_hpreq_ops;
2561         } else if (LDLM_CONVERT == lustre_msg_get_opc(req->rq_reqmsg)) {
2562                 req_capsule_set(&req->rq_pill, &RQF_LDLM_CONVERT);
2563                 req->rq_ops = &ldlm_cancel_hpreq_ops;
2564         }
2565         RETURN(0);
2566 }
2567
2568 static int ldlm_revoke_lock_cb(struct cfs_hash *hs, struct cfs_hash_bd *bd,
2569                                struct hlist_node *hnode, void *data)
2570
2571 {
2572         struct list_head *rpc_list = data;
2573         struct ldlm_lock *lock = cfs_hash_object(hs, hnode);
2574
2575         lock_res_and_lock(lock);
2576
2577         if (!ldlm_is_granted(lock)) {
2578                 unlock_res_and_lock(lock);
2579                 return 0;
2580         }
2581
2582         LASSERT(lock->l_resource);
2583         if (lock->l_resource->lr_type != LDLM_IBITS &&
2584             lock->l_resource->lr_type != LDLM_PLAIN) {
2585                 unlock_res_and_lock(lock);
2586                 return 0;
2587         }
2588
2589         if (ldlm_is_ast_sent(lock)) {
2590                 unlock_res_and_lock(lock);
2591                 return 0;
2592         }
2593
2594         LASSERT(lock->l_blocking_ast);
2595         LASSERT(!lock->l_blocking_lock);
2596
2597         ldlm_set_ast_sent(lock);
2598         if (lock->l_export && lock->l_export->exp_lock_hash) {
2599                 /*
2600                  * NB: it's safe to call cfs_hash_del() even lock isn't
2601                  * in exp_lock_hash.
2602                  */
2603                 /*
2604                  * In the function below, .hs_keycmp resolves to
2605                  * ldlm_export_lock_keycmp()
2606                  */
2607                 /* coverity[overrun-buffer-val] */
2608                 cfs_hash_del(lock->l_export->exp_lock_hash,
2609                              &lock->l_remote_handle, &lock->l_exp_hash);
2610         }
2611
2612         list_add_tail(&lock->l_rk_ast, rpc_list);
2613         LDLM_LOCK_GET(lock);
2614
2615         unlock_res_and_lock(lock);
2616         return 0;
2617 }
2618
2619 void ldlm_revoke_export_locks(struct obd_export *exp)
2620 {
2621         struct list_head rpc_list;
2622
2623         ENTRY;
2624
2625         INIT_LIST_HEAD(&rpc_list);
2626         cfs_hash_for_each_nolock(exp->exp_lock_hash,
2627                                  ldlm_revoke_lock_cb, &rpc_list, 0);
2628         ldlm_run_ast_work(exp->exp_obd->obd_namespace, &rpc_list,
2629                           LDLM_WORK_REVOKE_AST);
2630
2631         EXIT;
2632 }
2633 EXPORT_SYMBOL(ldlm_revoke_export_locks);
2634 #endif /* HAVE_SERVER_SUPPORT */
2635
2636 static int ldlm_bl_get_work(struct ldlm_bl_pool *blp,
2637                             struct ldlm_bl_work_item **p_blwi,
2638                             struct obd_export **p_exp)
2639 {
2640         struct ldlm_bl_work_item *blwi = NULL;
2641         static unsigned int num_bl;
2642         static unsigned int num_stale;
2643         int num_th = atomic_read(&blp->blp_num_threads);
2644
2645         *p_exp = obd_stale_export_get();
2646
2647         spin_lock(&blp->blp_lock);
2648         if (*p_exp != NULL) {
2649                 if (num_th == 1 || ++num_stale < num_th) {
2650                         spin_unlock(&blp->blp_lock);
2651                         return 1;
2652                 }
2653                 num_stale = 0;
2654         }
2655
2656         /* process a request from the blp_list at least every blp_num_threads */
2657         if (!list_empty(&blp->blp_list) &&
2658             (list_empty(&blp->blp_prio_list) || num_bl == 0))
2659                 blwi = list_entry(blp->blp_list.next,
2660                                   struct ldlm_bl_work_item, blwi_entry);
2661         else
2662                 if (!list_empty(&blp->blp_prio_list))
2663                         blwi = list_entry(blp->blp_prio_list.next,
2664                                           struct ldlm_bl_work_item,
2665                                           blwi_entry);
2666
2667         if (blwi) {
2668                 if (++num_bl >= num_th)
2669                         num_bl = 0;
2670                 list_del(&blwi->blwi_entry);
2671         }
2672         spin_unlock(&blp->blp_lock);
2673         *p_blwi = blwi;
2674
2675         if (*p_exp != NULL && *p_blwi != NULL) {
2676                 obd_stale_export_put(*p_exp);
2677                 *p_exp = NULL;
2678         }
2679
2680         return (*p_blwi != NULL || *p_exp != NULL) ? 1 : 0;
2681 }
2682
2683 /* This only contains temporary data until the thread starts */
2684 struct ldlm_bl_thread_data {
2685         struct ldlm_bl_pool     *bltd_blp;
2686         struct completion       bltd_comp;
2687         int                     bltd_num;
2688 };
2689
2690 static int ldlm_bl_thread_main(void *arg);
2691
2692 static int ldlm_bl_thread_start(struct ldlm_bl_pool *blp, bool check_busy)
2693 {
2694         struct ldlm_bl_thread_data bltd = { .bltd_blp = blp };
2695         struct task_struct *task;
2696
2697         init_completion(&bltd.bltd_comp);
2698
2699         bltd.bltd_num = atomic_inc_return(&blp->blp_num_threads);
2700         if (bltd.bltd_num >= blp->blp_max_threads) {
2701                 atomic_dec(&blp->blp_num_threads);
2702                 return 0;
2703         }
2704
2705         LASSERTF(bltd.bltd_num > 0, "thread num:%d\n", bltd.bltd_num);
2706         if (check_busy &&
2707             atomic_read(&blp->blp_busy_threads) < (bltd.bltd_num - 1)) {
2708                 atomic_dec(&blp->blp_num_threads);
2709                 return 0;
2710         }
2711
2712         task = kthread_run(ldlm_bl_thread_main, &bltd, "ldlm_bl_%02d",
2713                            bltd.bltd_num);
2714         if (IS_ERR(task)) {
2715                 CERROR("cannot start LDLM thread ldlm_bl_%02d: rc %ld\n",
2716                        bltd.bltd_num, PTR_ERR(task));
2717                 atomic_dec(&blp->blp_num_threads);
2718                 return PTR_ERR(task);
2719         }
2720         wait_for_completion(&bltd.bltd_comp);
2721
2722         return 0;
2723 }
2724
2725 /* Not fatal if racy and have a few too many threads */
2726 static int ldlm_bl_thread_need_create(struct ldlm_bl_pool *blp,
2727                                       struct ldlm_bl_work_item *blwi)
2728 {
2729         if (atomic_read(&blp->blp_num_threads) >= blp->blp_max_threads)
2730                 return 0;
2731
2732         if (atomic_read(&blp->blp_busy_threads) <
2733             atomic_read(&blp->blp_num_threads))
2734                 return 0;
2735
2736         if (blwi != NULL && (blwi->blwi_ns == NULL ||
2737                              blwi->blwi_mem_pressure))
2738                 return 0;
2739
2740         return 1;
2741 }
2742
2743 static int ldlm_bl_thread_blwi(struct ldlm_bl_pool *blp,
2744                                struct ldlm_bl_work_item *blwi)
2745 {
2746         ENTRY;
2747
2748         if (blwi->blwi_ns == NULL)
2749                 /* added by ldlm_cleanup() */
2750                 RETURN(LDLM_ITER_STOP);
2751
2752         if (blwi->blwi_mem_pressure)
2753                 memory_pressure_set();
2754
2755         OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_PAUSE_CANCEL2, 4);
2756
2757         if (blwi->blwi_count) {
2758                 int count;
2759                 /*
2760                  * The special case when we cancel locks in lru
2761                  * asynchronously, we pass the list of locks here.
2762                  * Thus locks are marked LDLM_FL_CANCELING, but NOT
2763                  * canceled locally yet.
2764                  */
2765                 count = ldlm_cli_cancel_list_local(&blwi->blwi_head,
2766                                                    blwi->blwi_count,
2767                                                    LCF_BL_AST);
2768                 ldlm_cli_cancel_list(&blwi->blwi_head, count, NULL,
2769                                      blwi->blwi_flags);
2770         } else {
2771                 ldlm_handle_bl_callback(blwi->blwi_ns, &blwi->blwi_ld,
2772                                         blwi->blwi_lock);
2773         }
2774         if (blwi->blwi_mem_pressure)
2775                 memory_pressure_clr();
2776
2777         if (blwi->blwi_flags & LCF_ASYNC)
2778                 OBD_FREE(blwi, sizeof(*blwi));
2779         else
2780                 complete(&blwi->blwi_comp);
2781
2782         RETURN(0);
2783 }
2784
2785 /**
2786  * Cancel stale locks on export. Cancel blocked locks first.
2787  * If the given export has blocked locks, the next in the list may have
2788  * them too, thus cancel not blocked locks only if the current export has
2789  * no blocked locks.
2790  **/
2791 static int ldlm_bl_thread_exports(struct ldlm_bl_pool *blp,
2792                                   struct obd_export *exp)
2793 {
2794         int num;
2795
2796         ENTRY;
2797
2798         OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_BL_EVICT, 4);
2799
2800         num = ldlm_export_cancel_blocked_locks(exp);
2801         if (num == 0)
2802                 ldlm_export_cancel_locks(exp);
2803
2804         obd_stale_export_put(exp);
2805
2806         RETURN(0);
2807 }
2808
2809
2810 /**
2811  * Main blocking requests processing thread.
2812  *
2813  * Callers put locks into its queue by calling ldlm_bl_to_thread.
2814  * This thread in the end ends up doing actual call to ->l_blocking_ast
2815  * for queued locks.
2816  */
2817 static int ldlm_bl_thread_main(void *arg)
2818 {
2819         struct lu_env *env;
2820         struct ldlm_bl_pool *blp;
2821         struct ldlm_bl_thread_data *bltd = arg;
2822         int rc;
2823
2824         ENTRY;
2825
2826         OBD_ALLOC_PTR(env);
2827         if (!env)
2828                 RETURN(-ENOMEM);
2829         rc = lu_env_init(env, LCT_DT_THREAD);
2830         if (rc)
2831                 GOTO(out_env, rc);
2832         rc = lu_env_add(env);
2833         if (rc)
2834                 GOTO(out_env_fini, rc);
2835
2836         blp = bltd->bltd_blp;
2837
2838         complete(&bltd->bltd_comp);
2839         /* cannot use bltd after this, it is only on caller's stack */
2840
2841         while (1) {
2842                 struct l_wait_info lwi = { 0 };
2843                 struct ldlm_bl_work_item *blwi = NULL;
2844                 struct obd_export *exp = NULL;
2845                 int rc;
2846
2847                 rc = ldlm_bl_get_work(blp, &blwi, &exp);
2848
2849                 if (rc == 0)
2850                         l_wait_event_exclusive(blp->blp_waitq,
2851                                                ldlm_bl_get_work(blp, &blwi,
2852                                                                 &exp),
2853                                                &lwi);
2854                 atomic_inc(&blp->blp_busy_threads);
2855
2856                 if (ldlm_bl_thread_need_create(blp, blwi))
2857                         /* discard the return value, we tried */
2858                         ldlm_bl_thread_start(blp, true);
2859
2860                 if (exp)
2861                         rc = ldlm_bl_thread_exports(blp, exp);
2862                 else if (blwi)
2863                         rc = ldlm_bl_thread_blwi(blp, blwi);
2864
2865                 atomic_dec(&blp->blp_busy_threads);
2866
2867                 if (rc == LDLM_ITER_STOP)
2868                         break;
2869
2870                 /*
2871                  * If there are many namespaces, we will not sleep waiting for
2872                  * work, and must do a cond_resched to avoid holding the CPU
2873                  * for too long
2874                  */
2875                 cond_resched();
2876         }
2877
2878         atomic_dec(&blp->blp_num_threads);
2879         complete(&blp->blp_comp);
2880
2881         lu_env_remove(env);
2882 out_env_fini:
2883         lu_env_fini(env);
2884 out_env:
2885         OBD_FREE_PTR(env);
2886         RETURN(rc);
2887 }
2888
2889
2890 static int ldlm_setup(void);
2891 static int ldlm_cleanup(void);
2892
2893 int ldlm_get_ref(void)
2894 {
2895         int rc = 0;
2896
2897         ENTRY;
2898         mutex_lock(&ldlm_ref_mutex);
2899         if (++ldlm_refcount == 1) {
2900                 rc = ldlm_setup();
2901                 if (rc)
2902                         ldlm_refcount--;
2903         }
2904         mutex_unlock(&ldlm_ref_mutex);
2905
2906         RETURN(rc);
2907 }
2908
2909 void ldlm_put_ref(void)
2910 {
2911         ENTRY;
2912         mutex_lock(&ldlm_ref_mutex);
2913         if (ldlm_refcount == 1) {
2914                 int rc = ldlm_cleanup();
2915
2916                 if (rc)
2917                         CERROR("ldlm_cleanup failed: %d\n", rc);
2918                 else
2919                         ldlm_refcount--;
2920         } else {
2921                 ldlm_refcount--;
2922         }
2923         mutex_unlock(&ldlm_ref_mutex);
2924
2925         EXIT;
2926 }
2927
2928 /*
2929  * Export handle<->lock hash operations.
2930  */
2931 static unsigned
2932 ldlm_export_lock_hash(struct cfs_hash *hs, const void *key, unsigned int mask)
2933 {
2934         return cfs_hash_u64_hash(((struct lustre_handle *)key)->cookie, mask);
2935 }
2936
2937 static void *
2938 ldlm_export_lock_key(struct hlist_node *hnode)
2939 {
2940         struct ldlm_lock *lock;
2941
2942         lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2943         return &lock->l_remote_handle;
2944 }
2945
2946 static void
2947 ldlm_export_lock_keycpy(struct hlist_node *hnode, void *key)
2948 {
2949         struct ldlm_lock     *lock;
2950
2951         lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2952         lock->l_remote_handle = *(struct lustre_handle *)key;
2953 }
2954
2955 static int
2956 ldlm_export_lock_keycmp(const void *key, struct hlist_node *hnode)
2957 {
2958         return lustre_handle_equal(ldlm_export_lock_key(hnode), key);
2959 }
2960
2961 static void *
2962 ldlm_export_lock_object(struct hlist_node *hnode)
2963 {
2964         return hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2965 }
2966
2967 static void
2968 ldlm_export_lock_get(struct cfs_hash *hs, struct hlist_node *hnode)
2969 {
2970         struct ldlm_lock *lock;
2971
2972         lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2973         LDLM_LOCK_GET(lock);
2974 }
2975
2976 static void
2977 ldlm_export_lock_put(struct cfs_hash *hs, struct hlist_node *hnode)
2978 {
2979         struct ldlm_lock *lock;
2980
2981         lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2982         LDLM_LOCK_RELEASE(lock);
2983 }
2984
2985 static struct cfs_hash_ops ldlm_export_lock_ops = {
2986         .hs_hash        = ldlm_export_lock_hash,
2987         .hs_key         = ldlm_export_lock_key,
2988         .hs_keycmp      = ldlm_export_lock_keycmp,
2989         .hs_keycpy      = ldlm_export_lock_keycpy,
2990         .hs_object      = ldlm_export_lock_object,
2991         .hs_get         = ldlm_export_lock_get,
2992         .hs_put         = ldlm_export_lock_put,
2993         .hs_put_locked  = ldlm_export_lock_put,
2994 };
2995
2996 int ldlm_init_export(struct obd_export *exp)
2997 {
2998         int rc;
2999
3000         ENTRY;
3001
3002         exp->exp_lock_hash =
3003                 cfs_hash_create(obd_uuid2str(&exp->exp_client_uuid),
3004                                 HASH_EXP_LOCK_CUR_BITS,
3005                                 HASH_EXP_LOCK_MAX_BITS,
3006                                 HASH_EXP_LOCK_BKT_BITS, 0,
3007                                 CFS_HASH_MIN_THETA, CFS_HASH_MAX_THETA,
3008                                 &ldlm_export_lock_ops,
3009                                 CFS_HASH_DEFAULT | CFS_HASH_REHASH_KEY |
3010                                 CFS_HASH_NBLK_CHANGE);
3011
3012         if (!exp->exp_lock_hash)
3013                 RETURN(-ENOMEM);
3014
3015         rc = ldlm_init_flock_export(exp);
3016         if (rc)
3017                 GOTO(err, rc);
3018
3019         RETURN(0);
3020 err:
3021         ldlm_destroy_export(exp);
3022         RETURN(rc);
3023 }
3024 EXPORT_SYMBOL(ldlm_init_export);
3025
3026 void ldlm_destroy_export(struct obd_export *exp)
3027 {
3028         ENTRY;
3029         cfs_hash_putref(exp->exp_lock_hash);
3030         exp->exp_lock_hash = NULL;
3031
3032         ldlm_destroy_flock_export(exp);
3033         EXIT;
3034 }
3035 EXPORT_SYMBOL(ldlm_destroy_export);
3036
3037 static ssize_t cancel_unused_locks_before_replay_show(struct kobject *kobj,
3038                                                       struct attribute *attr,
3039                                                       char *buf)
3040 {
3041         return sprintf(buf, "%d\n", ldlm_cancel_unused_locks_before_replay);
3042 }
3043
3044 static ssize_t cancel_unused_locks_before_replay_store(struct kobject *kobj,
3045                                                        struct attribute *attr,
3046                                                        const char *buffer,
3047                                                        size_t count)
3048 {
3049         int rc;
3050         unsigned long val;
3051
3052         rc = kstrtoul(buffer, 10, &val);
3053         if (rc)
3054                 return rc;
3055
3056         ldlm_cancel_unused_locks_before_replay = val;
3057
3058         return count;
3059 }
3060 LUSTRE_RW_ATTR(cancel_unused_locks_before_replay);
3061
3062 static struct attribute *ldlm_attrs[] = {
3063         &lustre_attr_cancel_unused_locks_before_replay.attr,
3064         NULL,
3065 };
3066
3067 static struct attribute_group ldlm_attr_group = {
3068         .attrs = ldlm_attrs,
3069 };
3070
3071 static int ldlm_setup(void)
3072 {
3073         static struct ptlrpc_service_conf       conf;
3074         struct ldlm_bl_pool                    *blp = NULL;
3075 #ifdef HAVE_SERVER_SUPPORT
3076         struct task_struct *task;
3077 #endif /* HAVE_SERVER_SUPPORT */
3078         int i;
3079         int rc = 0;
3080
3081         ENTRY;
3082
3083         if (ldlm_state != NULL)
3084                 RETURN(-EALREADY);
3085
3086         OBD_ALLOC(ldlm_state, sizeof(*ldlm_state));
3087         if (ldlm_state == NULL)
3088                 RETURN(-ENOMEM);
3089
3090         ldlm_kobj = kobject_create_and_add("ldlm", &lustre_kset->kobj);
3091         if (!ldlm_kobj)
3092                 GOTO(out, -ENOMEM);
3093
3094         rc = sysfs_create_group(ldlm_kobj, &ldlm_attr_group);
3095         if (rc)
3096                 GOTO(out, rc);
3097
3098         ldlm_ns_kset = kset_create_and_add("namespaces", NULL, ldlm_kobj);
3099         if (!ldlm_ns_kset)
3100                 GOTO(out, -ENOMEM);
3101
3102         ldlm_svc_kset = kset_create_and_add("services", NULL, ldlm_kobj);
3103         if (!ldlm_svc_kset)
3104                 GOTO(out, -ENOMEM);
3105
3106         rc = ldlm_debugfs_setup();
3107         if (rc != 0)
3108                 GOTO(out, rc);
3109
3110         memset(&conf, 0, sizeof(conf));
3111         conf = (typeof(conf)) {
3112                 .psc_name               = "ldlm_cbd",
3113                 .psc_watchdog_factor    = 2,
3114                 .psc_buf                = {
3115                         .bc_nbufs               = LDLM_CLIENT_NBUFS,
3116                         .bc_buf_size            = LDLM_BUFSIZE,
3117                         .bc_req_max_size        = LDLM_MAXREQSIZE,
3118                         .bc_rep_max_size        = LDLM_MAXREPSIZE,
3119                         .bc_req_portal          = LDLM_CB_REQUEST_PORTAL,
3120                         .bc_rep_portal          = LDLM_CB_REPLY_PORTAL,
3121                 },
3122                 .psc_thr                = {
3123                         .tc_thr_name            = "ldlm_cb",
3124                         .tc_thr_factor          = LDLM_THR_FACTOR,
3125                         .tc_nthrs_init          = LDLM_NTHRS_INIT,
3126                         .tc_nthrs_base          = LDLM_NTHRS_BASE,
3127                         .tc_nthrs_max           = LDLM_NTHRS_MAX,
3128                         .tc_nthrs_user          = ldlm_num_threads,
3129                         .tc_cpu_bind            = ldlm_cpu_bind,
3130                         .tc_ctx_tags            = LCT_MD_THREAD | LCT_DT_THREAD,
3131                 },
3132                 .psc_cpt                = {
3133                         .cc_pattern             = ldlm_cpts,
3134                         .cc_affinity            = true,
3135                 },
3136                 .psc_ops                = {
3137                         .so_req_handler         = ldlm_callback_handler,
3138                 },
3139         };
3140         ldlm_state->ldlm_cb_service = \
3141                         ptlrpc_register_service(&conf, ldlm_svc_kset,
3142                                                 ldlm_svc_debugfs_dir);
3143         if (IS_ERR(ldlm_state->ldlm_cb_service)) {
3144                 CERROR("failed to start service\n");
3145                 rc = PTR_ERR(ldlm_state->ldlm_cb_service);
3146                 ldlm_state->ldlm_cb_service = NULL;
3147                 GOTO(out, rc);
3148         }
3149
3150 #ifdef HAVE_SERVER_SUPPORT
3151         memset(&conf, 0, sizeof(conf));
3152         conf = (typeof(conf)) {
3153                 .psc_name               = "ldlm_canceld",
3154                 .psc_watchdog_factor    = 6,
3155                 .psc_buf                = {
3156                         .bc_nbufs               = LDLM_SERVER_NBUFS,
3157                         .bc_buf_size            = LDLM_BUFSIZE,
3158                         .bc_req_max_size        = LDLM_MAXREQSIZE,
3159                         .bc_rep_max_size        = LDLM_MAXREPSIZE,
3160                         .bc_req_portal          = LDLM_CANCEL_REQUEST_PORTAL,
3161                         .bc_rep_portal          = LDLM_CANCEL_REPLY_PORTAL,
3162
3163                 },
3164                 .psc_thr                = {
3165                         .tc_thr_name            = "ldlm_cn",
3166                         .tc_thr_factor          = LDLM_THR_FACTOR,
3167                         .tc_nthrs_init          = LDLM_NTHRS_INIT,
3168                         .tc_nthrs_base          = LDLM_NTHRS_BASE,
3169                         .tc_nthrs_max           = LDLM_NTHRS_MAX,
3170                         .tc_nthrs_user          = ldlm_num_threads,
3171                         .tc_cpu_bind            = ldlm_cpu_bind,
3172                         .tc_ctx_tags            = LCT_MD_THREAD | \
3173                                                   LCT_DT_THREAD | \
3174                                                   LCT_CL_THREAD,
3175                 },
3176                 .psc_cpt                = {
3177                         .cc_pattern             = ldlm_cpts,
3178                         .cc_affinity            = true,
3179                 },
3180                 .psc_ops                = {
3181                         .so_req_handler         = ldlm_cancel_handler,
3182                         .so_hpreq_handler       = ldlm_hpreq_handler,
3183                 },
3184         };
3185         ldlm_state->ldlm_cancel_service = \
3186                         ptlrpc_register_service(&conf, ldlm_svc_kset,
3187                                                 ldlm_svc_debugfs_dir);
3188         if (IS_ERR(ldlm_state->ldlm_cancel_service)) {
3189                 CERROR("failed to start service\n");
3190                 rc = PTR_ERR(ldlm_state->ldlm_cancel_service);
3191                 ldlm_state->ldlm_cancel_service = NULL;
3192                 GOTO(out, rc);
3193         }
3194 #endif /* HAVE_SERVER_SUPPORT */
3195
3196         OBD_ALLOC(blp, sizeof(*blp));
3197         if (blp == NULL)
3198                 GOTO(out, rc = -ENOMEM);
3199         ldlm_state->ldlm_bl_pool = blp;
3200
3201         spin_lock_init(&blp->blp_lock);
3202         INIT_LIST_HEAD(&blp->blp_list);
3203         INIT_LIST_HEAD(&blp->blp_prio_list);