Whamcloud - gitweb
fb7fc3bd41f16b8576bc90f613ffdcb29dbcad8f
[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.sun.com/software/products/lustre/docs/GPLv2.pdf
19  *
20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21  * CA 95054 USA or visit www.sun.com if you need additional information or
22  * have any questions.
23  *
24  * GPL HEADER END
25  */
26 /*
27  * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
28  * Use is subject to license terms.
29  *
30  * Copyright (c) 2010, 2013, Intel Corporation.
31  */
32 /*
33  * This file is part of Lustre, http://www.lustre.org/
34  * Lustre is a trademark of Sun Microsystems, Inc.
35  *
36  * lustre/ldlm/ldlm_lockd.c
37  *
38  * Author: Peter Braam <braam@clusterfs.com>
39  * Author: Phil Schwan <phil@clusterfs.com>
40  */
41
42 #define DEBUG_SUBSYSTEM S_LDLM
43
44 #ifdef __KERNEL__
45 # include <libcfs/libcfs.h>
46 #else
47 # include <liblustre.h>
48 #endif
49
50 #include <lustre_dlm.h>
51 #include <obd_class.h>
52 #include <libcfs/list.h>
53 #include "ldlm_internal.h"
54
55 static int ldlm_num_threads;
56 CFS_MODULE_PARM(ldlm_num_threads, "i", int, 0444,
57                 "number of DLM service threads to start");
58
59 static char *ldlm_cpts;
60 CFS_MODULE_PARM(ldlm_cpts, "s", charp, 0444,
61                 "CPU partitions ldlm threads should run on");
62
63 extern struct kmem_cache *ldlm_resource_slab;
64 extern struct kmem_cache *ldlm_lock_slab;
65 static struct mutex     ldlm_ref_mutex;
66 static int ldlm_refcount;
67
68 struct ldlm_cb_async_args {
69         struct ldlm_cb_set_arg *ca_set_arg;
70         struct ldlm_lock       *ca_lock;
71 };
72
73 /* LDLM state */
74
75 static struct ldlm_state *ldlm_state;
76
77 inline cfs_time_t round_timeout(cfs_time_t timeout)
78 {
79         return cfs_time_seconds((int)cfs_duration_sec(cfs_time_sub(timeout, 0)) + 1);
80 }
81
82 /* timeout for initial callback (AST) reply (bz10399) */
83 static inline unsigned int ldlm_get_rq_timeout(void)
84 {
85         /* Non-AT value */
86         unsigned int timeout = min(ldlm_timeout, obd_timeout / 3);
87
88         return timeout < 1 ? 1 : timeout;
89 }
90
91 #define ELT_STOPPED   0
92 #define ELT_READY     1
93 #define ELT_TERMINATE 2
94
95 struct ldlm_bl_pool {
96         spinlock_t              blp_lock;
97
98         /*
99          * blp_prio_list is used for callbacks that should be handled
100          * as a priority. It is used for LDLM_FL_DISCARD_DATA requests.
101          * see bug 13843
102          */
103         cfs_list_t              blp_prio_list;
104
105         /*
106          * blp_list is used for all other callbacks which are likely
107          * to take longer to process.
108          */
109         cfs_list_t              blp_list;
110
111         wait_queue_head_t       blp_waitq;
112         struct completion       blp_comp;
113         atomic_t            blp_num_threads;
114         atomic_t            blp_busy_threads;
115         int                     blp_min_threads;
116         int                     blp_max_threads;
117 };
118
119 struct ldlm_bl_work_item {
120         cfs_list_t              blwi_entry;
121         struct ldlm_namespace  *blwi_ns;
122         struct ldlm_lock_desc   blwi_ld;
123         struct ldlm_lock       *blwi_lock;
124         cfs_list_t              blwi_head;
125         int                     blwi_count;
126         struct completion        blwi_comp;
127         ldlm_cancel_flags_t     blwi_flags;
128         int                     blwi_mem_pressure;
129 };
130
131 #if defined(HAVE_SERVER_SUPPORT) && defined(__KERNEL__)
132
133 /**
134  * Protects both waiting_locks_list and expired_lock_thread.
135  */
136 static spinlock_t waiting_locks_spinlock;   /* BH lock (timer) */
137
138 /**
139  * List for contended locks.
140  *
141  * As soon as a lock is contended, it gets placed on this list and
142  * expected time to get a response is filled in the lock. A special
143  * thread walks the list looking for locks that should be released and
144  * schedules client evictions for those that have not been released in
145  * time.
146  *
147  * All access to it should be under waiting_locks_spinlock.
148  */
149 static cfs_list_t waiting_locks_list;
150 static struct timer_list waiting_locks_timer;
151
152 static struct expired_lock_thread {
153         wait_queue_head_t       elt_waitq;
154         int                     elt_state;
155         int                     elt_dump;
156         cfs_list_t              elt_expired_locks;
157 } expired_lock_thread;
158
159 static inline int have_expired_locks(void)
160 {
161         int need_to_run;
162
163         ENTRY;
164         spin_lock_bh(&waiting_locks_spinlock);
165         need_to_run = !cfs_list_empty(&expired_lock_thread.elt_expired_locks);
166         spin_unlock_bh(&waiting_locks_spinlock);
167
168         RETURN(need_to_run);
169 }
170
171 /**
172  * Check expired lock list for expired locks and time them out.
173  */
174 static int expired_lock_main(void *arg)
175 {
176         cfs_list_t *expired = &expired_lock_thread.elt_expired_locks;
177         struct l_wait_info lwi = { 0 };
178         int do_dump;
179
180         ENTRY;
181
182         expired_lock_thread.elt_state = ELT_READY;
183         wake_up(&expired_lock_thread.elt_waitq);
184
185         while (1) {
186                 l_wait_event(expired_lock_thread.elt_waitq,
187                              have_expired_locks() ||
188                              expired_lock_thread.elt_state == ELT_TERMINATE,
189                              &lwi);
190
191                 spin_lock_bh(&waiting_locks_spinlock);
192                 if (expired_lock_thread.elt_dump) {
193                         struct libcfs_debug_msg_data msgdata = {
194                                 .msg_file = __FILE__,
195                                 .msg_fn = "waiting_locks_callback",
196                                 .msg_line = expired_lock_thread.elt_dump };
197                         spin_unlock_bh(&waiting_locks_spinlock);
198
199                         /* from waiting_locks_callback, but not in timer */
200                         libcfs_debug_dumplog();
201                         libcfs_run_lbug_upcall(&msgdata);
202
203                         spin_lock_bh(&waiting_locks_spinlock);
204                         expired_lock_thread.elt_dump = 0;
205                 }
206
207                 do_dump = 0;
208
209                 while (!cfs_list_empty(expired)) {
210                         struct obd_export *export;
211                         struct ldlm_lock *lock;
212
213                         lock = cfs_list_entry(expired->next, struct ldlm_lock,
214                                           l_pending_chain);
215                         if ((void *)lock < LP_POISON + PAGE_CACHE_SIZE &&
216                             (void *)lock >= LP_POISON) {
217                                 spin_unlock_bh(&waiting_locks_spinlock);
218                                 CERROR("free lock on elt list %p\n", lock);
219                                 LBUG();
220                         }
221                         cfs_list_del_init(&lock->l_pending_chain);
222                         if ((void *)lock->l_export <
223                              LP_POISON + PAGE_CACHE_SIZE &&
224                             (void *)lock->l_export >= LP_POISON) {
225                                 CERROR("lock with free export on elt list %p\n",
226                                        lock->l_export);
227                                 lock->l_export = NULL;
228                                 LDLM_ERROR(lock, "free export");
229                                 /* release extra ref grabbed by
230                                  * ldlm_add_waiting_lock() or
231                                  * ldlm_failed_ast() */
232                                 LDLM_LOCK_RELEASE(lock);
233                                 continue;
234                         }
235
236                         if (ldlm_is_destroyed(lock)) {
237                                 /* release the lock refcount where
238                                  * waiting_locks_callback() founds */
239                                 LDLM_LOCK_RELEASE(lock);
240                                 continue;
241                         }
242                         export = class_export_lock_get(lock->l_export, lock);
243                         spin_unlock_bh(&waiting_locks_spinlock);
244
245                         do_dump++;
246                         class_fail_export(export);
247                         class_export_lock_put(export, lock);
248
249                         /* release extra ref grabbed by ldlm_add_waiting_lock()
250                          * or ldlm_failed_ast() */
251                         LDLM_LOCK_RELEASE(lock);
252
253                         spin_lock_bh(&waiting_locks_spinlock);
254                 }
255                 spin_unlock_bh(&waiting_locks_spinlock);
256
257                 if (do_dump && obd_dump_on_eviction) {
258                         CERROR("dump the log upon eviction\n");
259                         libcfs_debug_dumplog();
260                 }
261
262                 if (expired_lock_thread.elt_state == ELT_TERMINATE)
263                         break;
264         }
265
266         expired_lock_thread.elt_state = ELT_STOPPED;
267         wake_up(&expired_lock_thread.elt_waitq);
268         RETURN(0);
269 }
270
271 static int ldlm_add_waiting_lock(struct ldlm_lock *lock);
272 static int __ldlm_add_waiting_lock(struct ldlm_lock *lock, int seconds);
273
274 /**
275  * Check if there is a request in the export request list
276  * which prevents the lock canceling.
277  */
278 static int ldlm_lock_busy(struct ldlm_lock *lock)
279 {
280         struct ptlrpc_request *req;
281         int match = 0;
282         ENTRY;
283
284         if (lock->l_export == NULL)
285                 return 0;
286
287         spin_lock_bh(&lock->l_export->exp_rpc_lock);
288         cfs_list_for_each_entry(req, &lock->l_export->exp_hp_rpcs,
289                                 rq_exp_list) {
290                 if (req->rq_ops->hpreq_lock_match) {
291                         match = req->rq_ops->hpreq_lock_match(req, lock);
292                         if (match)
293                                 break;
294                 }
295         }
296         spin_unlock_bh(&lock->l_export->exp_rpc_lock);
297         RETURN(match);
298 }
299
300 /* This is called from within a timer interrupt and cannot schedule */
301 static void waiting_locks_callback(unsigned long unused)
302 {
303         struct ldlm_lock        *lock;
304         int                     need_dump = 0;
305
306         spin_lock_bh(&waiting_locks_spinlock);
307         while (!cfs_list_empty(&waiting_locks_list)) {
308                 lock = cfs_list_entry(waiting_locks_list.next, struct ldlm_lock,
309                                       l_pending_chain);
310                 if (cfs_time_after(lock->l_callback_timeout,
311                                    cfs_time_current()) ||
312                     (lock->l_req_mode == LCK_GROUP))
313                         break;
314
315                 /* Check if we need to prolong timeout */
316                 if (!OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_TIMEOUT) &&
317                     ldlm_lock_busy(lock)) {
318                         int cont = 1;
319
320                         if (lock->l_pending_chain.next == &waiting_locks_list)
321                                 cont = 0;
322
323                         LDLM_LOCK_GET(lock);
324
325                         spin_unlock_bh(&waiting_locks_spinlock);
326                         LDLM_DEBUG(lock, "prolong the busy lock");
327                         ldlm_refresh_waiting_lock(lock,
328                                                   ldlm_get_enq_timeout(lock));
329                         spin_lock_bh(&waiting_locks_spinlock);
330
331                         if (!cont) {
332                                 LDLM_LOCK_RELEASE(lock);
333                                 break;
334                         }
335
336                         LDLM_LOCK_RELEASE(lock);
337                         continue;
338                 }
339                 ldlm_lock_to_ns(lock)->ns_timeouts++;
340                 LDLM_ERROR(lock, "lock callback timer expired after %lds: "
341                            "evicting client at %s ",
342                            cfs_time_current_sec() - lock->l_last_activity,
343                            libcfs_nid2str(
344                                    lock->l_export->exp_connection->c_peer.nid));
345
346                 /* no needs to take an extra ref on the lock since it was in
347                  * the waiting_locks_list and ldlm_add_waiting_lock()
348                  * already grabbed a ref */
349                 cfs_list_del(&lock->l_pending_chain);
350                 cfs_list_add(&lock->l_pending_chain,
351                              &expired_lock_thread.elt_expired_locks);
352                 need_dump = 1;
353         }
354
355         if (!cfs_list_empty(&expired_lock_thread.elt_expired_locks)) {
356                 if (obd_dump_on_timeout && need_dump)
357                         expired_lock_thread.elt_dump = __LINE__;
358
359                 wake_up(&expired_lock_thread.elt_waitq);
360         }
361
362         /*
363          * Make sure the timer will fire again if we have any locks
364          * left.
365          */
366         if (!cfs_list_empty(&waiting_locks_list)) {
367                 cfs_time_t timeout_rounded;
368                 lock = cfs_list_entry(waiting_locks_list.next, struct ldlm_lock,
369                                       l_pending_chain);
370                 timeout_rounded = (cfs_time_t)round_timeout(lock->l_callback_timeout);
371                 cfs_timer_arm(&waiting_locks_timer, timeout_rounded);
372         }
373         spin_unlock_bh(&waiting_locks_spinlock);
374 }
375
376 /**
377  * Add lock to the list of contended locks.
378  *
379  * Indicate that we're waiting for a client to call us back cancelling a given
380  * lock.  We add it to the pending-callback chain, and schedule the lock-timeout
381  * timer to fire appropriately.  (We round up to the next second, to avoid
382  * floods of timer firings during periods of high lock contention and traffic).
383  * As done by ldlm_add_waiting_lock(), the caller must grab a lock reference
384  * if it has been added to the waiting list (1 is returned).
385  *
386  * Called with the namespace lock held.
387  */
388 static int __ldlm_add_waiting_lock(struct ldlm_lock *lock, int seconds)
389 {
390         cfs_time_t timeout;
391         cfs_time_t timeout_rounded;
392
393         if (!cfs_list_empty(&lock->l_pending_chain))
394                 return 0;
395
396         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_NOTIMEOUT) ||
397             OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_TIMEOUT))
398                 seconds = 1;
399
400         timeout = cfs_time_shift(seconds);
401         if (likely(cfs_time_after(timeout, lock->l_callback_timeout)))
402                 lock->l_callback_timeout = timeout;
403
404         timeout_rounded = round_timeout(lock->l_callback_timeout);
405
406         if (cfs_time_before(timeout_rounded,
407                             cfs_timer_deadline(&waiting_locks_timer)) ||
408             !cfs_timer_is_armed(&waiting_locks_timer)) {
409                 cfs_timer_arm(&waiting_locks_timer, timeout_rounded);
410         }
411         /* if the new lock has a shorter timeout than something earlier on
412            the list, we'll wait the longer amount of time; no big deal. */
413         /* FIFO */
414         cfs_list_add_tail(&lock->l_pending_chain, &waiting_locks_list);
415         return 1;
416 }
417
418 static int ldlm_add_waiting_lock(struct ldlm_lock *lock)
419 {
420         int ret;
421         int timeout = ldlm_get_enq_timeout(lock);
422
423         /* NB: must be called with hold of lock_res_and_lock() */
424         LASSERT(ldlm_is_res_locked(lock));
425         ldlm_set_waited(lock);
426
427         LASSERT(!ldlm_is_cancel_on_block(lock));
428
429         spin_lock_bh(&waiting_locks_spinlock);
430         if (ldlm_is_destroyed(lock)) {
431                 static cfs_time_t next;
432                 spin_unlock_bh(&waiting_locks_spinlock);
433                 LDLM_ERROR(lock, "not waiting on destroyed lock (bug 5653)");
434                 if (cfs_time_after(cfs_time_current(), next)) {
435                         next = cfs_time_shift(14400);
436                         libcfs_debug_dumpstack(NULL);
437                 }
438                 return 0;
439         }
440
441         ret = __ldlm_add_waiting_lock(lock, timeout);
442         if (ret) {
443                 /* grab ref on the lock if it has been added to the
444                  * waiting list */
445                 LDLM_LOCK_GET(lock);
446         }
447         spin_unlock_bh(&waiting_locks_spinlock);
448
449         if (ret) {
450                 spin_lock_bh(&lock->l_export->exp_bl_list_lock);
451                 if (cfs_list_empty(&lock->l_exp_list))
452                         cfs_list_add(&lock->l_exp_list,
453                                      &lock->l_export->exp_bl_list);
454                 spin_unlock_bh(&lock->l_export->exp_bl_list_lock);
455         }
456
457         LDLM_DEBUG(lock, "%sadding to wait list(timeout: %d, AT: %s)",
458                    ret == 0 ? "not re-" : "", timeout,
459                    AT_OFF ? "off" : "on");
460         return ret;
461 }
462
463 /**
464  * Remove a lock from the pending list, likely because it had its cancellation
465  * callback arrive without incident.  This adjusts the lock-timeout timer if
466  * needed.  Returns 0 if the lock wasn't pending after all, 1 if it was.
467  * As done by ldlm_del_waiting_lock(), the caller must release the lock
468  * reference when the lock is removed from any list (1 is returned).
469  *
470  * Called with namespace lock held.
471  */
472 static int __ldlm_del_waiting_lock(struct ldlm_lock *lock)
473 {
474         cfs_list_t *list_next;
475
476         if (cfs_list_empty(&lock->l_pending_chain))
477                 return 0;
478
479         list_next = lock->l_pending_chain.next;
480         if (lock->l_pending_chain.prev == &waiting_locks_list) {
481                 /* Removing the head of the list, adjust timer. */
482                 if (list_next == &waiting_locks_list) {
483                         /* No more, just cancel. */
484                         cfs_timer_disarm(&waiting_locks_timer);
485                 } else {
486                         struct ldlm_lock *next;
487                         next = cfs_list_entry(list_next, struct ldlm_lock,
488                                               l_pending_chain);
489                         cfs_timer_arm(&waiting_locks_timer,
490                                       round_timeout(next->l_callback_timeout));
491                 }
492         }
493         cfs_list_del_init(&lock->l_pending_chain);
494
495         return 1;
496 }
497
498 int ldlm_del_waiting_lock(struct ldlm_lock *lock)
499 {
500         int ret;
501
502         if (lock->l_export == NULL) {
503                 /* We don't have a "waiting locks list" on clients. */
504                 CDEBUG(D_DLMTRACE, "Client lock %p : no-op\n", lock);
505                 return 0;
506         }
507
508         spin_lock_bh(&waiting_locks_spinlock);
509         ret = __ldlm_del_waiting_lock(lock);
510         spin_unlock_bh(&waiting_locks_spinlock);
511
512         /* remove the lock out of export blocking list */
513         spin_lock_bh(&lock->l_export->exp_bl_list_lock);
514         cfs_list_del_init(&lock->l_exp_list);
515         spin_unlock_bh(&lock->l_export->exp_bl_list_lock);
516
517         if (ret) {
518                 /* release lock ref if it has indeed been removed
519                  * from a list */
520                 LDLM_LOCK_RELEASE(lock);
521         }
522
523         LDLM_DEBUG(lock, "%s", ret == 0 ? "wasn't waiting" : "removed");
524         return ret;
525 }
526 EXPORT_SYMBOL(ldlm_del_waiting_lock);
527
528 /**
529  * Prolong the contended lock waiting time.
530  *
531  * Called with namespace lock held.
532  */
533 int ldlm_refresh_waiting_lock(struct ldlm_lock *lock, int timeout)
534 {
535         if (lock->l_export == NULL) {
536                 /* We don't have a "waiting locks list" on clients. */
537                 LDLM_DEBUG(lock, "client lock: no-op");
538                 return 0;
539         }
540
541         spin_lock_bh(&waiting_locks_spinlock);
542
543         if (cfs_list_empty(&lock->l_pending_chain)) {
544                 spin_unlock_bh(&waiting_locks_spinlock);
545                 LDLM_DEBUG(lock, "wasn't waiting");
546                 return 0;
547         }
548
549         /* we remove/add the lock to the waiting list, so no needs to
550          * release/take a lock reference */
551         __ldlm_del_waiting_lock(lock);
552         __ldlm_add_waiting_lock(lock, timeout);
553         spin_unlock_bh(&waiting_locks_spinlock);
554
555         LDLM_DEBUG(lock, "refreshed");
556         return 1;
557 }
558 EXPORT_SYMBOL(ldlm_refresh_waiting_lock);
559
560 #else /* !HAVE_SERVER_SUPPORT ||  !__KERNEL__ */
561
562 int ldlm_del_waiting_lock(struct ldlm_lock *lock)
563 {
564         RETURN(0);
565 }
566
567 int ldlm_refresh_waiting_lock(struct ldlm_lock *lock, int timeout)
568 {
569         RETURN(0);
570 }
571
572 # ifdef HAVE_SERVER_SUPPORT
573 static int ldlm_add_waiting_lock(struct ldlm_lock *lock)
574 {
575         LASSERT(ldlm_is_res_locked(lock) && !ldlm_is_cancel_on_block(lock));
576         RETURN(1);
577 }
578
579 # endif
580 #endif /* HAVE_SERVER_SUPPORT && __KERNEL__ */
581
582 #ifdef HAVE_SERVER_SUPPORT
583
584 /**
585  * Perform lock cleanup if AST sending failed.
586  */
587 static void ldlm_failed_ast(struct ldlm_lock *lock, int rc,
588                             const char *ast_type)
589 {
590         LCONSOLE_ERROR_MSG(0x138, "%s: A client on nid %s was evicted due "
591                            "to a lock %s callback time out: rc %d\n",
592                            lock->l_export->exp_obd->obd_name,
593                            obd_export_nid2str(lock->l_export), ast_type, rc);
594
595         if (obd_dump_on_timeout)
596                 libcfs_debug_dumplog();
597 #ifdef __KERNEL__
598         spin_lock_bh(&waiting_locks_spinlock);
599         if (__ldlm_del_waiting_lock(lock) == 0)
600                 /* the lock was not in any list, grab an extra ref before adding
601                  * the lock to the expired list */
602                 LDLM_LOCK_GET(lock);
603         cfs_list_add(&lock->l_pending_chain,
604                      &expired_lock_thread.elt_expired_locks);
605         wake_up(&expired_lock_thread.elt_waitq);
606         spin_unlock_bh(&waiting_locks_spinlock);
607 #else
608         class_fail_export(lock->l_export);
609 #endif
610 }
611
612 /**
613  * Perform lock cleanup if AST reply came with error.
614  */
615 static int ldlm_handle_ast_error(struct ldlm_lock *lock,
616                                  struct ptlrpc_request *req, int rc,
617                                  const char *ast_type)
618 {
619         lnet_process_id_t peer = req->rq_import->imp_connection->c_peer;
620
621         if (rc == -ETIMEDOUT || rc == -EINTR || rc == -ENOTCONN) {
622                 LASSERT(lock->l_export);
623                 if (lock->l_export->exp_libclient) {
624                         LDLM_DEBUG(lock, "%s AST to liblustre client (nid %s)"
625                                    " timeout, just cancelling lock", ast_type,
626                                    libcfs_nid2str(peer.nid));
627                         ldlm_lock_cancel(lock);
628                         rc = -ERESTART;
629                 } else if (ldlm_is_cancel(lock)) {
630                         LDLM_DEBUG(lock, "%s AST timeout from nid %s, but "
631                                    "cancel was received (AST reply lost?)",
632                                    ast_type, libcfs_nid2str(peer.nid));
633                         ldlm_lock_cancel(lock);
634                         rc = -ERESTART;
635                 } else {
636                         ldlm_del_waiting_lock(lock);
637                         ldlm_failed_ast(lock, rc, ast_type);
638                 }
639         } else if (rc) {
640                 if (rc == -EINVAL) {
641                         struct ldlm_resource *res = lock->l_resource;
642                         LDLM_DEBUG(lock, "client (nid %s) returned %d"
643                                " from %s AST - normal race",
644                                libcfs_nid2str(peer.nid),
645                                req->rq_repmsg ?
646                                lustre_msg_get_status(req->rq_repmsg) : -1,
647                                ast_type);
648                         if (res) {
649                                 /* update lvbo to return proper attributes.
650                                  * see bug 23174 */
651                                 ldlm_resource_getref(res);
652                                 ldlm_res_lvbo_update(res, NULL, 1);
653                                 ldlm_resource_putref(res);
654                         }
655
656                 } else {
657                         LDLM_ERROR(lock, "client (nid %s) returned %d "
658                                    "from %s AST", libcfs_nid2str(peer.nid),
659                                    (req->rq_repmsg != NULL) ?
660                                    lustre_msg_get_status(req->rq_repmsg) : 0,
661                                    ast_type);
662                 }
663                 ldlm_lock_cancel(lock);
664                 /* Server-side AST functions are called from ldlm_reprocess_all,
665                  * which needs to be told to please restart its reprocessing. */
666                 rc = -ERESTART;
667         }
668
669         return rc;
670 }
671
672 static int ldlm_cb_interpret(const struct lu_env *env,
673                              struct ptlrpc_request *req, void *data, int rc)
674 {
675         struct ldlm_cb_async_args *ca   = data;
676         struct ldlm_lock          *lock = ca->ca_lock;
677         struct ldlm_cb_set_arg    *arg  = ca->ca_set_arg;
678         ENTRY;
679
680         LASSERT(lock != NULL);
681
682         switch (arg->type) {
683         case LDLM_GL_CALLBACK:
684                 /* Update the LVB from disk if the AST failed
685                  * (this is a legal race)
686                  *
687                  * - Glimpse callback of local lock just returns
688                  *   -ELDLM_NO_LOCK_DATA.
689                  * - Glimpse callback of remote lock might return
690                  *   -ELDLM_NO_LOCK_DATA when inode is cleared. LU-274
691                  */
692                 if (rc == -ELDLM_NO_LOCK_DATA) {
693                         LDLM_DEBUG(lock, "lost race - client has a lock but no "
694                                    "inode");
695                         ldlm_res_lvbo_update(lock->l_resource, NULL, 1);
696                 } else if (rc != 0) {
697                         rc = ldlm_handle_ast_error(lock, req, rc, "glimpse");
698                 } else {
699                         rc = ldlm_res_lvbo_update(lock->l_resource, req, 1);
700                 }
701                 break;
702         case LDLM_BL_CALLBACK:
703                 if (rc != 0)
704                         rc = ldlm_handle_ast_error(lock, req, rc, "blocking");
705                 break;
706         case LDLM_CP_CALLBACK:
707                 if (rc != 0)
708                         rc = ldlm_handle_ast_error(lock, req, rc, "completion");
709                 break;
710         default:
711                 LDLM_ERROR(lock, "invalid opcode for lock callback %d",
712                            arg->type);
713                 LBUG();
714         }
715
716         /* release extra reference taken in ldlm_ast_fini() */
717         LDLM_LOCK_RELEASE(lock);
718
719         if (rc == -ERESTART)
720                 atomic_inc(&arg->restart);
721
722         RETURN(0);
723 }
724
725 static inline int ldlm_ast_fini(struct ptlrpc_request *req,
726                                 struct ldlm_cb_set_arg *arg,
727                                 struct ldlm_lock *lock,
728                                 int instant_cancel)
729 {
730         int rc = 0;
731         ENTRY;
732
733         if (unlikely(instant_cancel)) {
734                 rc = ptl_send_rpc(req, 1);
735                 ptlrpc_req_finished(req);
736                 if (rc == 0)
737                         atomic_inc(&arg->restart);
738         } else {
739                 LDLM_LOCK_GET(lock);
740                 ptlrpc_set_add_req(arg->set, req);
741         }
742
743         RETURN(rc);
744 }
745
746 /**
747  * Check if there are requests in the export request list which prevent
748  * the lock canceling and make these requests high priority ones.
749  */
750 static void ldlm_lock_reorder_req(struct ldlm_lock *lock)
751 {
752         struct ptlrpc_request *req;
753         ENTRY;
754
755         if (lock->l_export == NULL) {
756                 LDLM_DEBUG(lock, "client lock: no-op");
757                 RETURN_EXIT;
758         }
759
760         spin_lock_bh(&lock->l_export->exp_rpc_lock);
761         cfs_list_for_each_entry(req, &lock->l_export->exp_hp_rpcs,
762                                 rq_exp_list) {
763                 /* Do not process requests that were not yet added to there
764                  * incoming queue or were already removed from there for
765                  * processing. We evaluate ptlrpc_nrs_req_can_move() without
766                  * holding svcpt->scp_req_lock, and then redo the check with
767                  * the lock held once we need to obtain a reliable result.
768                  */
769                 if (ptlrpc_nrs_req_can_move(req) &&
770                     req->rq_ops->hpreq_lock_match &&
771                     req->rq_ops->hpreq_lock_match(req, lock))
772                         ptlrpc_nrs_req_hp_move(req);
773         }
774         spin_unlock_bh(&lock->l_export->exp_rpc_lock);
775         EXIT;
776 }
777
778 /**
779  * ->l_blocking_ast() method for server-side locks. This is invoked when newly
780  * enqueued server lock conflicts with given one.
781  *
782  * Sends blocking AST RPC to the client owning that lock; arms timeout timer
783  * to wait for client response.
784  */
785 int ldlm_server_blocking_ast(struct ldlm_lock *lock,
786                              struct ldlm_lock_desc *desc,
787                              void *data, int flag)
788 {
789         struct ldlm_cb_async_args *ca;
790         struct ldlm_cb_set_arg *arg = data;
791         struct ldlm_request    *body;
792         struct ptlrpc_request  *req;
793         int                     instant_cancel = 0;
794         int                     rc = 0;
795         ENTRY;
796
797         if (flag == LDLM_CB_CANCELING)
798                 /* Don't need to do anything here. */
799                 RETURN(0);
800
801         LASSERT(lock);
802         LASSERT(data != NULL);
803         if (lock->l_export->exp_obd->obd_recovering != 0)
804                 LDLM_ERROR(lock, "BUG 6063: lock collide during recovery");
805
806         ldlm_lock_reorder_req(lock);
807
808         req = ptlrpc_request_alloc_pack(lock->l_export->exp_imp_reverse,
809                                         &RQF_LDLM_BL_CALLBACK,
810                                         LUSTRE_DLM_VERSION, LDLM_BL_CALLBACK);
811         if (req == NULL)
812                 RETURN(-ENOMEM);
813
814         CLASSERT(sizeof(*ca) <= sizeof(req->rq_async_args));
815         ca = ptlrpc_req_async_args(req);
816         ca->ca_set_arg = arg;
817         ca->ca_lock = lock;
818
819         req->rq_interpret_reply = ldlm_cb_interpret;
820         req->rq_no_resend = 1;
821
822         lock_res_and_lock(lock);
823         if (lock->l_granted_mode != lock->l_req_mode) {
824                 /* this blocking AST will be communicated as part of the
825                  * completion AST instead */
826                 unlock_res_and_lock(lock);
827
828                 ptlrpc_req_finished(req);
829                 LDLM_DEBUG(lock, "lock not granted, not sending blocking AST");
830                 RETURN(0);
831         }
832
833         if (ldlm_is_destroyed(lock)) {
834                 /* What's the point? */
835                 unlock_res_and_lock(lock);
836                 ptlrpc_req_finished(req);
837                 RETURN(0);
838         }
839
840         if (ldlm_is_cancel_on_block(lock))
841                 instant_cancel = 1;
842
843         body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
844         body->lock_handle[0] = lock->l_remote_handle;
845         body->lock_desc = *desc;
846         body->lock_flags |= ldlm_flags_to_wire(lock->l_flags & LDLM_FL_AST_MASK);
847
848         LDLM_DEBUG(lock, "server preparing blocking AST");
849
850         ptlrpc_request_set_replen(req);
851         if (instant_cancel) {
852                 unlock_res_and_lock(lock);
853                 ldlm_lock_cancel(lock);
854         } else {
855                 LASSERT(lock->l_granted_mode == lock->l_req_mode);
856                 ldlm_add_waiting_lock(lock);
857                 unlock_res_and_lock(lock);
858         }
859
860         req->rq_send_state = LUSTRE_IMP_FULL;
861         /* ptlrpc_request_alloc_pack already set timeout */
862         if (AT_OFF)
863                 req->rq_timeout = ldlm_get_rq_timeout();
864
865         lock->l_last_activity = cfs_time_current_sec();
866
867         if (lock->l_export && lock->l_export->exp_nid_stats &&
868             lock->l_export->exp_nid_stats->nid_ldlm_stats)
869                 lprocfs_counter_incr(lock->l_export->exp_nid_stats->nid_ldlm_stats,
870                                      LDLM_BL_CALLBACK - LDLM_FIRST_OPC);
871
872         rc = ldlm_ast_fini(req, arg, lock, instant_cancel);
873
874         RETURN(rc);
875 }
876 EXPORT_SYMBOL(ldlm_server_blocking_ast);
877
878 /**
879  * ->l_completion_ast callback for a remote lock in server namespace.
880  *
881  *  Sends AST to the client notifying it of lock granting.  If initial
882  *  lock response was not sent yet, instead of sending another RPC, just
883  *  mark the lock as granted and client will understand
884  */
885 int ldlm_server_completion_ast(struct ldlm_lock *lock, __u64 flags, void *data)
886 {
887         struct ldlm_cb_set_arg *arg = data;
888         struct ldlm_request    *body;
889         struct ptlrpc_request  *req;
890         struct ldlm_cb_async_args *ca;
891         long                    total_enqueue_wait;
892         int                     instant_cancel = 0;
893         int                     rc = 0;
894         int                     lvb_len;
895         ENTRY;
896
897         LASSERT(lock != NULL);
898         LASSERT(data != NULL);
899
900         total_enqueue_wait = cfs_time_sub(cfs_time_current_sec(),
901                                           lock->l_last_activity);
902
903         if (OBD_FAIL_PRECHECK(OBD_FAIL_OST_LDLM_REPLY_NET)) {
904                 LDLM_DEBUG(lock, "dropping CP AST");
905                 RETURN(0);
906         }
907
908         req = ptlrpc_request_alloc(lock->l_export->exp_imp_reverse,
909                                     &RQF_LDLM_CP_CALLBACK);
910         if (req == NULL)
911                 RETURN(-ENOMEM);
912
913         /* server namespace, doesn't need lock */
914         lvb_len = ldlm_lvbo_size(lock);
915         /* LU-3124 & LU-2187: to not return layout in completion AST because
916          * it may deadlock for LU-2187, or client may not have enough space
917          * for large layout. The layout will be returned to client with an
918          * extra RPC to fetch xattr.lov */
919         if (ldlm_has_layout(lock))
920                 lvb_len = 0;
921
922         req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_CLIENT, lvb_len);
923         rc = ptlrpc_request_pack(req, LUSTRE_DLM_VERSION, LDLM_CP_CALLBACK);
924         if (rc) {
925                 ptlrpc_request_free(req);
926                 RETURN(rc);
927         }
928
929         CLASSERT(sizeof(*ca) <= sizeof(req->rq_async_args));
930         ca = ptlrpc_req_async_args(req);
931         ca->ca_set_arg = arg;
932         ca->ca_lock = lock;
933
934         req->rq_interpret_reply = ldlm_cb_interpret;
935         req->rq_no_resend = 1;
936         body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
937
938         body->lock_handle[0] = lock->l_remote_handle;
939         body->lock_flags = ldlm_flags_to_wire(flags);
940         ldlm_lock2desc(lock, &body->lock_desc);
941         if (lvb_len > 0) {
942                 void *lvb = req_capsule_client_get(&req->rq_pill, &RMF_DLM_LVB);
943
944                 lvb_len = ldlm_lvbo_fill(lock, lvb, lvb_len);
945                 if (lvb_len < 0) {
946                         /* We still need to send the RPC to wake up the blocked
947                          * enqueue thread on the client.
948                          *
949                          * Consider old client, there is no better way to notify
950                          * the failure, just zero-sized the LVB, then the client
951                          * will fail out as "-EPROTO". */
952                         req_capsule_shrink(&req->rq_pill, &RMF_DLM_LVB, 0,
953                                            RCL_CLIENT);
954                         instant_cancel = 1;
955                 } else {
956                         req_capsule_shrink(&req->rq_pill, &RMF_DLM_LVB, lvb_len,
957                                            RCL_CLIENT);
958                 }
959         }
960
961         LDLM_DEBUG(lock, "server preparing completion AST (after %lds wait)",
962                    total_enqueue_wait);
963
964         lock->l_last_activity = cfs_time_current_sec();
965
966         /* Server-side enqueue wait time estimate, used in
967             __ldlm_add_waiting_lock to set future enqueue timers */
968         if (total_enqueue_wait < ldlm_get_enq_timeout(lock))
969                 at_measured(ldlm_lock_to_ns_at(lock),
970                             total_enqueue_wait);
971         else
972                 /* bz18618. Don't add lock enqueue time we spend waiting for a
973                    previous callback to fail. Locks waiting legitimately will
974                    get extended by ldlm_refresh_waiting_lock regardless of the
975                    estimate, so it's okay to underestimate here. */
976                 LDLM_DEBUG(lock, "lock completed after %lus; estimate was %ds. "
977                        "It is likely that a previous callback timed out.",
978                        total_enqueue_wait,
979                        at_get(ldlm_lock_to_ns_at(lock)));
980
981         ptlrpc_request_set_replen(req);
982
983         req->rq_send_state = LUSTRE_IMP_FULL;
984         /* ptlrpc_request_pack already set timeout */
985         if (AT_OFF)
986                 req->rq_timeout = ldlm_get_rq_timeout();
987
988         /* We only send real blocking ASTs after the lock is granted */
989         lock_res_and_lock(lock);
990         if (ldlm_is_ast_sent(lock)) {
991                 body->lock_flags |= ldlm_flags_to_wire(LDLM_FL_AST_SENT);
992                 /* Copy AST flags like LDLM_FL_DISCARD_DATA. */
993                 body->lock_flags |= ldlm_flags_to_wire(lock->l_flags &
994                                                        LDLM_FL_AST_MASK);
995
996                 /* We might get here prior to ldlm_handle_enqueue setting
997                  * LDLM_FL_CANCEL_ON_BLOCK flag. Then we will put this lock
998                  * into waiting list, but this is safe and similar code in
999                  * ldlm_handle_enqueue will call ldlm_lock_cancel() still,
1000                  * that would not only cancel the lock, but will also remove
1001                  * it from waiting list */
1002                 if (ldlm_is_cancel_on_block(lock)) {
1003                         unlock_res_and_lock(lock);
1004                         ldlm_lock_cancel(lock);
1005                         instant_cancel = 1;
1006                         lock_res_and_lock(lock);
1007                 } else {
1008                         /* start the lock-timeout clock */
1009                         ldlm_add_waiting_lock(lock);
1010                 }
1011         }
1012         unlock_res_and_lock(lock);
1013
1014         if (lock->l_export && lock->l_export->exp_nid_stats &&
1015             lock->l_export->exp_nid_stats->nid_ldlm_stats)
1016                 lprocfs_counter_incr(lock->l_export->exp_nid_stats->nid_ldlm_stats,
1017                                      LDLM_CP_CALLBACK - LDLM_FIRST_OPC);
1018
1019         rc = ldlm_ast_fini(req, arg, lock, instant_cancel);
1020
1021         RETURN(lvb_len < 0 ? lvb_len : rc);
1022 }
1023 EXPORT_SYMBOL(ldlm_server_completion_ast);
1024
1025 /**
1026  * Server side ->l_glimpse_ast handler for client locks.
1027  *
1028  * Sends glimpse AST to the client and waits for reply. Then updates
1029  * lvbo with the result.
1030  */
1031 int ldlm_server_glimpse_ast(struct ldlm_lock *lock, void *data)
1032 {
1033         struct ldlm_cb_set_arg          *arg = data;
1034         struct ldlm_request             *body;
1035         struct ptlrpc_request           *req;
1036         struct ldlm_cb_async_args       *ca;
1037         int                              rc;
1038         struct req_format               *req_fmt;
1039         ENTRY;
1040
1041         LASSERT(lock != NULL);
1042
1043         if (arg->gl_desc != NULL)
1044                 /* There is a glimpse descriptor to pack */
1045                 req_fmt = &RQF_LDLM_GL_DESC_CALLBACK;
1046         else
1047                 req_fmt = &RQF_LDLM_GL_CALLBACK;
1048
1049         req = ptlrpc_request_alloc_pack(lock->l_export->exp_imp_reverse,
1050                                         req_fmt, LUSTRE_DLM_VERSION,
1051                                         LDLM_GL_CALLBACK);
1052
1053         if (req == NULL)
1054                 RETURN(-ENOMEM);
1055
1056         if (arg->gl_desc != NULL) {
1057                 /* copy the GL descriptor */
1058                 union ldlm_gl_desc      *desc;
1059                 desc = req_capsule_client_get(&req->rq_pill, &RMF_DLM_GL_DESC);
1060                 *desc = *arg->gl_desc;
1061         }
1062
1063         body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1064         body->lock_handle[0] = lock->l_remote_handle;
1065         ldlm_lock2desc(lock, &body->lock_desc);
1066
1067         CLASSERT(sizeof(*ca) <= sizeof(req->rq_async_args));
1068         ca = ptlrpc_req_async_args(req);
1069         ca->ca_set_arg = arg;
1070         ca->ca_lock = lock;
1071
1072         /* server namespace, doesn't need lock */
1073         req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
1074                              ldlm_lvbo_size(lock));
1075         ptlrpc_request_set_replen(req);
1076
1077         req->rq_send_state = LUSTRE_IMP_FULL;
1078         /* ptlrpc_request_alloc_pack already set timeout */
1079         if (AT_OFF)
1080                 req->rq_timeout = ldlm_get_rq_timeout();
1081
1082         lock->l_last_activity = cfs_time_current_sec();
1083
1084         req->rq_interpret_reply = ldlm_cb_interpret;
1085
1086         if (lock->l_export && lock->l_export->exp_nid_stats &&
1087             lock->l_export->exp_nid_stats->nid_ldlm_stats)
1088                 lprocfs_counter_incr(lock->l_export->exp_nid_stats->nid_ldlm_stats,
1089                                      LDLM_GL_CALLBACK - LDLM_FIRST_OPC);
1090
1091         rc = ldlm_ast_fini(req, arg, lock, 0);
1092
1093         RETURN(rc);
1094 }
1095 EXPORT_SYMBOL(ldlm_server_glimpse_ast);
1096
1097 int ldlm_glimpse_locks(struct ldlm_resource *res, cfs_list_t *gl_work_list)
1098 {
1099         int     rc;
1100         ENTRY;
1101
1102         rc = ldlm_run_ast_work(ldlm_res_to_ns(res), gl_work_list,
1103                                LDLM_WORK_GL_AST);
1104         if (rc == -ERESTART)
1105                 ldlm_reprocess_all(res);
1106
1107         RETURN(rc);
1108 }
1109 EXPORT_SYMBOL(ldlm_glimpse_locks);
1110
1111 /* return LDLM lock associated with a lock callback request */
1112 struct ldlm_lock *ldlm_request_lock(struct ptlrpc_request *req)
1113 {
1114         struct ldlm_cb_async_args       *ca;
1115         struct ldlm_lock                *lock;
1116         ENTRY;
1117
1118         ca = ptlrpc_req_async_args(req);
1119         lock = ca->ca_lock;
1120         if (lock == NULL)
1121                 RETURN(ERR_PTR(-EFAULT));
1122
1123         RETURN(lock);
1124 }
1125 EXPORT_SYMBOL(ldlm_request_lock);
1126
1127 static void ldlm_svc_get_eopc(const struct ldlm_request *dlm_req,
1128                        struct lprocfs_stats *srv_stats)
1129 {
1130         int lock_type = 0, op = 0;
1131
1132         lock_type = dlm_req->lock_desc.l_resource.lr_type;
1133
1134         switch (lock_type) {
1135         case LDLM_PLAIN:
1136                 op = PTLRPC_LAST_CNTR + LDLM_PLAIN_ENQUEUE;
1137                 break;
1138         case LDLM_EXTENT:
1139                 if (dlm_req->lock_flags & LDLM_FL_HAS_INTENT)
1140                         op = PTLRPC_LAST_CNTR + LDLM_GLIMPSE_ENQUEUE;
1141                 else
1142                         op = PTLRPC_LAST_CNTR + LDLM_EXTENT_ENQUEUE;
1143                 break;
1144         case LDLM_FLOCK:
1145                 op = PTLRPC_LAST_CNTR + LDLM_FLOCK_ENQUEUE;
1146                 break;
1147         case LDLM_IBITS:
1148                 op = PTLRPC_LAST_CNTR + LDLM_IBITS_ENQUEUE;
1149                 break;
1150         default:
1151                 op = 0;
1152                 break;
1153         }
1154
1155         if (op)
1156                 lprocfs_counter_incr(srv_stats, op);
1157
1158         return;
1159 }
1160
1161 /**
1162  * Main server-side entry point into LDLM for enqueue. This is called by ptlrpc
1163  * service threads to carry out client lock enqueueing requests.
1164  */
1165 int ldlm_handle_enqueue0(struct ldlm_namespace *ns,
1166                          struct ptlrpc_request *req,
1167                          const struct ldlm_request *dlm_req,
1168                          const struct ldlm_callback_suite *cbs)
1169 {
1170         struct ldlm_reply *dlm_rep;
1171         __u64 flags;
1172         ldlm_error_t err = ELDLM_OK;
1173         struct ldlm_lock *lock = NULL;
1174         void *cookie = NULL;
1175         int rc = 0;
1176         struct ldlm_resource *res = NULL;
1177         ENTRY;
1178
1179         LDLM_DEBUG_NOLOCK("server-side enqueue handler START");
1180
1181         ldlm_request_cancel(req, dlm_req, LDLM_ENQUEUE_CANCEL_OFF);
1182         flags = ldlm_flags_from_wire(dlm_req->lock_flags);
1183
1184         LASSERT(req->rq_export);
1185
1186         if (ptlrpc_req2svc(req)->srv_stats != NULL)
1187                 ldlm_svc_get_eopc(dlm_req, ptlrpc_req2svc(req)->srv_stats);
1188
1189         if (req->rq_export && req->rq_export->exp_nid_stats &&
1190             req->rq_export->exp_nid_stats->nid_ldlm_stats)
1191                 lprocfs_counter_incr(req->rq_export->exp_nid_stats->nid_ldlm_stats,
1192                                      LDLM_ENQUEUE - LDLM_FIRST_OPC);
1193
1194         if (unlikely(dlm_req->lock_desc.l_resource.lr_type < LDLM_MIN_TYPE ||
1195                      dlm_req->lock_desc.l_resource.lr_type >= LDLM_MAX_TYPE)) {
1196                 DEBUG_REQ(D_ERROR, req, "invalid lock request type %d",
1197                           dlm_req->lock_desc.l_resource.lr_type);
1198                 GOTO(out, rc = -EFAULT);
1199         }
1200
1201         if (unlikely(dlm_req->lock_desc.l_req_mode <= LCK_MINMODE ||
1202                      dlm_req->lock_desc.l_req_mode >= LCK_MAXMODE ||
1203                      dlm_req->lock_desc.l_req_mode &
1204                      (dlm_req->lock_desc.l_req_mode-1))) {
1205                 DEBUG_REQ(D_ERROR, req, "invalid lock request mode %d",
1206                           dlm_req->lock_desc.l_req_mode);
1207                 GOTO(out, rc = -EFAULT);
1208         }
1209
1210         if (exp_connect_flags(req->rq_export) & OBD_CONNECT_IBITS) {
1211                 if (unlikely(dlm_req->lock_desc.l_resource.lr_type ==
1212                              LDLM_PLAIN)) {
1213                         DEBUG_REQ(D_ERROR, req,
1214                                   "PLAIN lock request from IBITS client?");
1215                         GOTO(out, rc = -EPROTO);
1216                 }
1217         } else if (unlikely(dlm_req->lock_desc.l_resource.lr_type ==
1218                             LDLM_IBITS)) {
1219                 DEBUG_REQ(D_ERROR, req,
1220                           "IBITS lock request from unaware client?");
1221                 GOTO(out, rc = -EPROTO);
1222         }
1223
1224 #if 0
1225         /* FIXME this makes it impossible to use LDLM_PLAIN locks -- check
1226            against server's _CONNECT_SUPPORTED flags? (I don't want to use
1227            ibits for mgc/mgs) */
1228
1229         /* INODEBITS_INTEROP: Perform conversion from plain lock to
1230          * inodebits lock if client does not support them. */
1231         if (!(exp_connect_flags(req->rq_export) & OBD_CONNECT_IBITS) &&
1232             (dlm_req->lock_desc.l_resource.lr_type == LDLM_PLAIN)) {
1233                 dlm_req->lock_desc.l_resource.lr_type = LDLM_IBITS;
1234                 dlm_req->lock_desc.l_policy_data.l_inodebits.bits =
1235                         MDS_INODELOCK_LOOKUP | MDS_INODELOCK_UPDATE;
1236                 if (dlm_req->lock_desc.l_req_mode == LCK_PR)
1237                         dlm_req->lock_desc.l_req_mode = LCK_CR;
1238         }
1239 #endif
1240
1241         if (unlikely((flags & LDLM_FL_REPLAY) ||
1242                      (lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT))) {
1243                 /* Find an existing lock in the per-export lock hash */
1244                 /* In the function below, .hs_keycmp resolves to
1245                  * ldlm_export_lock_keycmp() */
1246                 /* coverity[overrun-buffer-val] */
1247                 lock = cfs_hash_lookup(req->rq_export->exp_lock_hash,
1248                                        (void *)&dlm_req->lock_handle[0]);
1249                 if (lock != NULL) {
1250                         DEBUG_REQ(D_DLMTRACE, req, "found existing lock cookie "
1251                                   LPX64, lock->l_handle.h_cookie);
1252                         flags |= LDLM_FL_RESENT;
1253                         GOTO(existing_lock, rc = 0);
1254                 }
1255         }
1256
1257         /* The lock's callback data might be set in the policy function */
1258         lock = ldlm_lock_create(ns, &dlm_req->lock_desc.l_resource.lr_name,
1259                                 dlm_req->lock_desc.l_resource.lr_type,
1260                                 dlm_req->lock_desc.l_req_mode,
1261                                 cbs, NULL, 0, LVB_T_NONE);
1262         if (IS_ERR(lock)) {
1263                 rc = PTR_ERR(lock);
1264                 lock = NULL;
1265                 GOTO(out, rc);
1266         }
1267
1268         lock->l_last_activity = cfs_time_current_sec();
1269         lock->l_remote_handle = dlm_req->lock_handle[0];
1270         LDLM_DEBUG(lock, "server-side enqueue handler, new lock created");
1271
1272         /* Initialize resource lvb but not for a lock being replayed since
1273          * Client already got lvb sent in this case.
1274          * This must occur early since some policy methods assume resource
1275          * lvb is available (lr_lvb_data != NULL).
1276          */
1277         res = lock->l_resource;
1278         if (!(flags & LDLM_FL_REPLAY)) {
1279                 /* non-replayed lock, delayed lvb init may need to be done */
1280                 rc = ldlm_lvbo_init(res);
1281                 if (rc < 0) {
1282                         LDLM_ERROR(lock, "delayed lvb init failed (rc %d)", rc);
1283                         GOTO(out, rc);
1284                 }
1285         }
1286
1287         OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_ENQUEUE_BLOCKED, obd_timeout * 2);
1288         /* Don't enqueue a lock onto the export if it is been disonnected
1289          * due to eviction (bug 3822) or server umount (bug 24324).
1290          * Cancel it now instead. */
1291         if (req->rq_export->exp_disconnected) {
1292                 LDLM_ERROR(lock, "lock on disconnected export %p",
1293                            req->rq_export);
1294                 GOTO(out, rc = -ENOTCONN);
1295         }
1296
1297         lock->l_export = class_export_lock_get(req->rq_export, lock);
1298         if (lock->l_export->exp_lock_hash)
1299                 cfs_hash_add(lock->l_export->exp_lock_hash,
1300                              &lock->l_remote_handle,
1301                              &lock->l_exp_hash);
1302
1303 existing_lock:
1304
1305         if (flags & LDLM_FL_HAS_INTENT) {
1306                 /* In this case, the reply buffer is allocated deep in
1307                  * local_lock_enqueue by the policy function. */
1308                 cookie = req;
1309         } else {
1310                 /* based on the assumption that lvb size never changes during
1311                  * resource life time otherwise it need resource->lr_lock's
1312                  * protection */
1313                 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB,
1314                                      RCL_SERVER, ldlm_lvbo_size(lock));
1315
1316                 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_ENQUEUE_EXTENT_ERR))
1317                         GOTO(out, rc = -ENOMEM);
1318
1319                 rc = req_capsule_server_pack(&req->rq_pill);
1320                 if (rc)
1321                         GOTO(out, rc);
1322         }
1323
1324         if (dlm_req->lock_desc.l_resource.lr_type != LDLM_PLAIN)
1325                 ldlm_convert_policy_to_local(req->rq_export,
1326                                           dlm_req->lock_desc.l_resource.lr_type,
1327                                           &dlm_req->lock_desc.l_policy_data,
1328                                           &lock->l_policy_data);
1329         if (dlm_req->lock_desc.l_resource.lr_type == LDLM_EXTENT)
1330                 lock->l_req_extent = lock->l_policy_data.l_extent;
1331
1332         err = ldlm_lock_enqueue(ns, &lock, cookie, &flags);
1333         if (err) {
1334                 if ((int)err < 0)
1335                         rc = (int)err;
1336                 GOTO(out, err);
1337         }
1338
1339         dlm_rep = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
1340         dlm_rep->lock_flags = ldlm_flags_to_wire(flags);
1341
1342         ldlm_lock2desc(lock, &dlm_rep->lock_desc);
1343         ldlm_lock2handle(lock, &dlm_rep->lock_handle);
1344
1345         /* We never send a blocking AST until the lock is granted, but
1346          * we can tell it right now */
1347         lock_res_and_lock(lock);
1348
1349         /* Now take into account flags to be inherited from original lock
1350            request both in reply to client and in our own lock flags. */
1351         dlm_rep->lock_flags |= dlm_req->lock_flags & LDLM_FL_INHERIT_MASK;
1352         lock->l_flags |= ldlm_flags_from_wire(dlm_req->lock_flags &
1353                                               LDLM_FL_INHERIT_MASK);
1354
1355         /* Don't move a pending lock onto the export if it has already been
1356          * disconnected due to eviction (bug 5683) or server umount (bug 24324).
1357          * Cancel it now instead. */
1358         if (unlikely(req->rq_export->exp_disconnected ||
1359                      OBD_FAIL_CHECK(OBD_FAIL_LDLM_ENQUEUE_OLD_EXPORT))) {
1360                 LDLM_ERROR(lock, "lock on destroyed export %p", req->rq_export);
1361                 rc = -ENOTCONN;
1362         } else if (ldlm_is_ast_sent(lock)) {
1363                 dlm_rep->lock_flags |= ldlm_flags_to_wire(LDLM_FL_AST_SENT);
1364                 if (lock->l_granted_mode == lock->l_req_mode) {
1365                         /*
1366                          * Only cancel lock if it was granted, because it would
1367                          * be destroyed immediately and would never be granted
1368                          * in the future, causing timeouts on client.  Not
1369                          * granted lock will be cancelled immediately after
1370                          * sending completion AST.
1371                          */
1372                         if (dlm_rep->lock_flags & LDLM_FL_CANCEL_ON_BLOCK) {
1373                                 unlock_res_and_lock(lock);
1374                                 ldlm_lock_cancel(lock);
1375                                 lock_res_and_lock(lock);
1376                         } else
1377                                 ldlm_add_waiting_lock(lock);
1378                 }
1379         }
1380         /* Make sure we never ever grant usual metadata locks to liblustre
1381            clients */
1382         if ((dlm_req->lock_desc.l_resource.lr_type == LDLM_PLAIN ||
1383             dlm_req->lock_desc.l_resource.lr_type == LDLM_IBITS) &&
1384              req->rq_export->exp_libclient) {
1385                 if (unlikely(!ldlm_is_cancel_on_block(lock) ||
1386                              !(dlm_rep->lock_flags & LDLM_FL_CANCEL_ON_BLOCK))){
1387                         CERROR("Granting sync lock to libclient. "
1388                                "req fl %d, rep fl %d, lock fl "LPX64"\n",
1389                                dlm_req->lock_flags, dlm_rep->lock_flags,
1390                                lock->l_flags);
1391                         LDLM_ERROR(lock, "sync lock");
1392                         if (dlm_req->lock_flags & LDLM_FL_HAS_INTENT) {
1393                                 struct ldlm_intent *it;
1394
1395                                 it = req_capsule_client_get(&req->rq_pill,
1396                                                             &RMF_LDLM_INTENT);
1397                                 if (it != NULL) {
1398                                         CERROR("This is intent %s ("LPU64")\n",
1399                                                ldlm_it2str(it->opc), it->opc);
1400                                 }
1401                         }
1402                 }
1403         }
1404
1405         unlock_res_and_lock(lock);
1406
1407         EXIT;
1408  out:
1409         req->rq_status = rc ?: err; /* return either error - bug 11190 */
1410         if (!req->rq_packed_final) {
1411                 err = lustre_pack_reply(req, 1, NULL, NULL);
1412                 if (rc == 0)
1413                         rc = err;
1414         }
1415
1416         /* The LOCK_CHANGED code in ldlm_lock_enqueue depends on this
1417          * ldlm_reprocess_all.  If this moves, revisit that code. -phil */
1418         if (lock != NULL) {
1419                 LDLM_DEBUG(lock, "server-side enqueue handler, sending reply"
1420                            "(err=%d, rc=%d)", err, rc);
1421
1422                 if (rc == 0) {
1423                         if (req_capsule_has_field(&req->rq_pill, &RMF_DLM_LVB,
1424                                                   RCL_SERVER) &&
1425                             ldlm_lvbo_size(lock) > 0) {
1426                                 void *buf;
1427                                 int buflen;
1428
1429                                 buf = req_capsule_server_get(&req->rq_pill,
1430                                                              &RMF_DLM_LVB);
1431                                 LASSERTF(buf != NULL, "req %p, lock %p\n",
1432                                          req, lock);
1433                                 buflen = req_capsule_get_size(&req->rq_pill,
1434                                                 &RMF_DLM_LVB, RCL_SERVER);
1435                                 /* non-replayed lock, delayed lvb init may
1436                                  * need to be occur now */
1437                                 if ((buflen > 0) && !(flags & LDLM_FL_REPLAY)) {
1438                                         buflen = ldlm_lvbo_fill(lock, buf,
1439                                                                 buflen);
1440                                         if (buflen >= 0)
1441                                                 req_capsule_shrink(
1442                                                         &req->rq_pill,
1443                                                         &RMF_DLM_LVB,
1444                                                         buflen, RCL_SERVER);
1445                                         else
1446                                                 rc = buflen;
1447                                 } else if (flags & LDLM_FL_REPLAY) {
1448                                         /* no LVB resend upon replay */
1449                                         if (buflen > 0)
1450                                                 req_capsule_shrink(
1451                                                         &req->rq_pill,
1452                                                         &RMF_DLM_LVB,
1453                                                         0, RCL_SERVER);
1454                                         else
1455                                                 rc = buflen;
1456                                 } else {
1457                                         rc = buflen;
1458                                 }
1459                         }
1460                 }
1461
1462                 if (rc != 0) {
1463                         lock_res_and_lock(lock);
1464                         ldlm_resource_unlink_lock(lock);
1465                         ldlm_lock_destroy_nolock(lock);
1466                         unlock_res_and_lock(lock);
1467                 }
1468
1469                 if (!err && dlm_req->lock_desc.l_resource.lr_type != LDLM_FLOCK)
1470                         ldlm_reprocess_all(lock->l_resource);
1471
1472                 LDLM_LOCK_RELEASE(lock);
1473         }
1474
1475         LDLM_DEBUG_NOLOCK("server-side enqueue handler END (lock %p, rc %d)",
1476                           lock, rc);
1477
1478         return rc;
1479 }
1480 EXPORT_SYMBOL(ldlm_handle_enqueue0);
1481
1482 /**
1483  * Old-style LDLM main entry point for server code enqueue.
1484  */
1485 int ldlm_handle_enqueue(struct ptlrpc_request *req,
1486                         ldlm_completion_callback completion_callback,
1487                         ldlm_blocking_callback blocking_callback,
1488                         ldlm_glimpse_callback glimpse_callback)
1489 {
1490         struct ldlm_request *dlm_req;
1491         struct ldlm_callback_suite cbs = {
1492                 .lcs_completion = completion_callback,
1493                 .lcs_blocking   = blocking_callback,
1494                 .lcs_glimpse    = glimpse_callback
1495         };
1496         int rc;
1497
1498         dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1499         if (dlm_req != NULL) {
1500                 rc = ldlm_handle_enqueue0(req->rq_export->exp_obd->obd_namespace,
1501                                           req, dlm_req, &cbs);
1502         } else {
1503                 rc = -EFAULT;
1504         }
1505         return rc;
1506 }
1507 EXPORT_SYMBOL(ldlm_handle_enqueue);
1508
1509 /**
1510  * Main LDLM entry point for server code to process lock conversion requests.
1511  */
1512 int ldlm_handle_convert0(struct ptlrpc_request *req,
1513                          const struct ldlm_request *dlm_req)
1514 {
1515         struct ldlm_reply *dlm_rep;
1516         struct ldlm_lock *lock;
1517         int rc;
1518         ENTRY;
1519
1520         if (req->rq_export && req->rq_export->exp_nid_stats &&
1521             req->rq_export->exp_nid_stats->nid_ldlm_stats)
1522                 lprocfs_counter_incr(req->rq_export->exp_nid_stats->nid_ldlm_stats,
1523                                      LDLM_CONVERT - LDLM_FIRST_OPC);
1524
1525         rc = req_capsule_server_pack(&req->rq_pill);
1526         if (rc)
1527                 RETURN(rc);
1528
1529         dlm_rep = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
1530         dlm_rep->lock_flags = dlm_req->lock_flags;
1531
1532         lock = ldlm_handle2lock(&dlm_req->lock_handle[0]);
1533         if (!lock) {
1534                 req->rq_status = LUSTRE_EINVAL;
1535         } else {
1536                 void *res = NULL;
1537
1538                 LDLM_DEBUG(lock, "server-side convert handler START");
1539
1540                 lock->l_last_activity = cfs_time_current_sec();
1541                 res = ldlm_lock_convert(lock, dlm_req->lock_desc.l_req_mode,
1542                                         &dlm_rep->lock_flags);
1543                 if (res) {
1544                         if (ldlm_del_waiting_lock(lock))
1545                                 LDLM_DEBUG(lock, "converted waiting lock");
1546                         req->rq_status = 0;
1547                 } else {
1548                         req->rq_status = LUSTRE_EDEADLK;
1549                 }
1550         }
1551
1552         if (lock) {
1553                 if (!req->rq_status)
1554                         ldlm_reprocess_all(lock->l_resource);
1555                 LDLM_DEBUG(lock, "server-side convert handler END");
1556                 LDLM_LOCK_PUT(lock);
1557         } else
1558                 LDLM_DEBUG_NOLOCK("server-side convert handler END");
1559
1560         RETURN(0);
1561 }
1562 EXPORT_SYMBOL(ldlm_handle_convert0);
1563
1564 /**
1565  * Old-style main LDLM entry point for server code to process lock conversion
1566  * requests.
1567  */
1568 int ldlm_handle_convert(struct ptlrpc_request *req)
1569 {
1570         int rc;
1571         struct ldlm_request *dlm_req;
1572
1573         dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1574         if (dlm_req != NULL) {
1575                 rc = ldlm_handle_convert0(req, dlm_req);
1576         } else {
1577                 CERROR ("Can't unpack dlm_req\n");
1578                 rc = -EFAULT;
1579         }
1580         return rc;
1581 }
1582 EXPORT_SYMBOL(ldlm_handle_convert);
1583
1584 /**
1585  * Cancel all the locks whose handles are packed into ldlm_request
1586  *
1587  * Called by server code expecting such combined cancel activity
1588  * requests.
1589  */
1590 int ldlm_request_cancel(struct ptlrpc_request *req,
1591                         const struct ldlm_request *dlm_req, int first)
1592 {
1593         struct ldlm_resource *res, *pres = NULL;
1594         struct ldlm_lock *lock;
1595         int i, count, done = 0;
1596         ENTRY;
1597
1598         count = dlm_req->lock_count ? dlm_req->lock_count : 1;
1599         if (first >= count)
1600                 RETURN(0);
1601
1602         if (count == 1 && dlm_req->lock_handle[0].cookie == 0)
1603                 RETURN(0);
1604
1605         /* There is no lock on the server at the replay time,
1606          * skip lock cancelling to make replay tests to pass. */
1607         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
1608                 RETURN(0);
1609
1610         LDLM_DEBUG_NOLOCK("server-side cancel handler START: %d locks, "
1611                           "starting at %d", count, first);
1612
1613         for (i = first; i < count; i++) {
1614                 lock = ldlm_handle2lock(&dlm_req->lock_handle[i]);
1615                 if (!lock) {
1616                         LDLM_DEBUG_NOLOCK("server-side cancel handler stale "
1617                                           "lock (cookie "LPU64")",
1618                                           dlm_req->lock_handle[i].cookie);
1619                         continue;
1620                 }
1621
1622                 res = lock->l_resource;
1623                 done++;
1624
1625                 /* This code is an optimization to only attempt lock
1626                  * granting on the resource (that could be CPU-expensive)
1627                  * after we are done cancelling lock in that resource. */
1628                 if (res != pres) {
1629                         if (pres != NULL) {
1630                                 ldlm_reprocess_all(pres);
1631                                 LDLM_RESOURCE_DELREF(pres);
1632                                 ldlm_resource_putref(pres);
1633                         }
1634                         if (res != NULL) {
1635                                 ldlm_resource_getref(res);
1636                                 LDLM_RESOURCE_ADDREF(res);
1637                                 ldlm_res_lvbo_update(res, NULL, 1);
1638                         }
1639                         pres = res;
1640                 }
1641                 ldlm_lock_cancel(lock);
1642                 LDLM_LOCK_PUT(lock);
1643         }
1644         if (pres != NULL) {
1645                 ldlm_reprocess_all(pres);
1646                 LDLM_RESOURCE_DELREF(pres);
1647                 ldlm_resource_putref(pres);
1648         }
1649         LDLM_DEBUG_NOLOCK("server-side cancel handler END");
1650         RETURN(done);
1651 }
1652 EXPORT_SYMBOL(ldlm_request_cancel);
1653
1654 /**
1655  * Main LDLM entry point for server code to cancel locks.
1656  *
1657  * Typically gets called from service handler on LDLM_CANCEL opc.
1658  */
1659 int ldlm_handle_cancel(struct ptlrpc_request *req)
1660 {
1661         struct ldlm_request *dlm_req;
1662         int rc;
1663         ENTRY;
1664
1665         dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1666         if (dlm_req == NULL) {
1667                 CDEBUG(D_INFO, "bad request buffer for cancel\n");
1668                 RETURN(-EFAULT);
1669         }
1670
1671         if (req->rq_export && req->rq_export->exp_nid_stats &&
1672             req->rq_export->exp_nid_stats->nid_ldlm_stats)
1673                 lprocfs_counter_incr(req->rq_export->exp_nid_stats->nid_ldlm_stats,
1674                                      LDLM_CANCEL - LDLM_FIRST_OPC);
1675
1676         rc = req_capsule_server_pack(&req->rq_pill);
1677         if (rc)
1678                 RETURN(rc);
1679
1680         if (!ldlm_request_cancel(req, dlm_req, 0))
1681                 req->rq_status = LUSTRE_ESTALE;
1682
1683         RETURN(ptlrpc_reply(req));
1684 }
1685 EXPORT_SYMBOL(ldlm_handle_cancel);
1686 #endif /* HAVE_SERVER_SUPPORT */
1687
1688 /**
1689  * Callback handler for receiving incoming blocking ASTs.
1690  *
1691  * This can only happen on client side.
1692  */
1693 void ldlm_handle_bl_callback(struct ldlm_namespace *ns,
1694                              struct ldlm_lock_desc *ld, struct ldlm_lock *lock)
1695 {
1696         int do_ast;
1697         ENTRY;
1698
1699         LDLM_DEBUG(lock, "client blocking AST callback handler");
1700
1701         lock_res_and_lock(lock);
1702         ldlm_set_cbpending(lock);
1703
1704         if (ldlm_is_cancel_on_block(lock))
1705                 ldlm_set_cancel(lock);
1706
1707         do_ast = (!lock->l_readers && !lock->l_writers);
1708         unlock_res_and_lock(lock);
1709
1710         if (do_ast) {
1711                 CDEBUG(D_DLMTRACE, "Lock %p already unused, calling callback (%p)\n",
1712                        lock, lock->l_blocking_ast);
1713                 if (lock->l_blocking_ast != NULL)
1714                         lock->l_blocking_ast(lock, ld, lock->l_ast_data,
1715                                              LDLM_CB_BLOCKING);
1716         } else {
1717                 CDEBUG(D_DLMTRACE, "Lock %p is referenced, will be cancelled later\n",
1718                        lock);
1719         }
1720
1721         LDLM_DEBUG(lock, "client blocking callback handler END");
1722         LDLM_LOCK_RELEASE(lock);
1723         EXIT;
1724 }
1725
1726 /**
1727  * Callback handler for receiving incoming completion ASTs.
1728  *
1729  * This only can happen on client side.
1730  */
1731 static void ldlm_handle_cp_callback(struct ptlrpc_request *req,
1732                                     struct ldlm_namespace *ns,
1733                                     struct ldlm_request *dlm_req,
1734                                     struct ldlm_lock *lock)
1735 {
1736         int lvb_len;
1737         CFS_LIST_HEAD(ast_list);
1738         int rc = 0;
1739         ENTRY;
1740
1741         LDLM_DEBUG(lock, "client completion callback handler START");
1742
1743         if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_BL_CB_RACE)) {
1744                 int to = cfs_time_seconds(1);
1745                 while (to > 0) {
1746                         schedule_timeout_and_set_state(
1747                                 TASK_INTERRUPTIBLE, to);
1748                         if (lock->l_granted_mode == lock->l_req_mode ||
1749                             ldlm_is_destroyed(lock))
1750                                 break;
1751                 }
1752         }
1753
1754         lvb_len = req_capsule_get_size(&req->rq_pill, &RMF_DLM_LVB, RCL_CLIENT);
1755         if (lvb_len < 0) {
1756                 LDLM_ERROR(lock, "Fail to get lvb_len, rc = %d", lvb_len);
1757                 GOTO(out, rc = lvb_len);
1758         } else if (lvb_len > 0) {
1759                 if (lock->l_lvb_len > 0) {
1760                         /* for extent lock, lvb contains ost_lvb{}. */
1761                         LASSERT(lock->l_lvb_data != NULL);
1762
1763                         if (unlikely(lock->l_lvb_len < lvb_len)) {
1764                                 LDLM_ERROR(lock, "Replied LVB is larger than "
1765                                            "expectation, expected = %d, "
1766                                            "replied = %d",
1767                                            lock->l_lvb_len, lvb_len);
1768                                 GOTO(out, rc = -EINVAL);
1769                         }
1770                 } else if (ldlm_has_layout(lock)) { /* for layout lock, lvb has
1771                                                      * variable length */
1772                         void *lvb_data;
1773
1774                         OBD_ALLOC_LARGE(lvb_data, lvb_len);
1775                         if (lvb_data == NULL) {
1776                                 LDLM_ERROR(lock, "No memory: %d.\n", lvb_len);
1777                                 GOTO(out, rc = -ENOMEM);
1778                         }
1779
1780                         lock_res_and_lock(lock);
1781                         LASSERT(lock->l_lvb_data == NULL);
1782                         lock->l_lvb_type = LVB_T_LAYOUT;
1783                         lock->l_lvb_data = lvb_data;
1784                         lock->l_lvb_len = lvb_len;
1785                         unlock_res_and_lock(lock);
1786                 }
1787         }
1788
1789         lock_res_and_lock(lock);
1790         if (ldlm_is_destroyed(lock) ||
1791             lock->l_granted_mode == lock->l_req_mode) {
1792                 /* bug 11300: the lock has already been granted */
1793                 unlock_res_and_lock(lock);
1794                 LDLM_DEBUG(lock, "Double grant race happened");
1795                 GOTO(out, rc = 0);
1796         }
1797
1798         /* If we receive the completion AST before the actual enqueue returned,
1799          * then we might need to switch lock modes, resources, or extents. */
1800         if (dlm_req->lock_desc.l_granted_mode != lock->l_req_mode) {
1801                 lock->l_req_mode = dlm_req->lock_desc.l_granted_mode;
1802                 LDLM_DEBUG(lock, "completion AST, new lock mode");
1803         }
1804
1805         if (lock->l_resource->lr_type != LDLM_PLAIN) {
1806                 ldlm_convert_policy_to_local(req->rq_export,
1807                                           dlm_req->lock_desc.l_resource.lr_type,
1808                                           &dlm_req->lock_desc.l_policy_data,
1809                                           &lock->l_policy_data);
1810                 LDLM_DEBUG(lock, "completion AST, new policy data");
1811         }
1812
1813         ldlm_resource_unlink_lock(lock);
1814         if (memcmp(&dlm_req->lock_desc.l_resource.lr_name,
1815                    &lock->l_resource->lr_name,
1816                    sizeof(lock->l_resource->lr_name)) != 0) {
1817                 unlock_res_and_lock(lock);
1818                 rc = ldlm_lock_change_resource(ns, lock,
1819                                 &dlm_req->lock_desc.l_resource.lr_name);
1820                 if (rc < 0) {
1821                         LDLM_ERROR(lock, "Failed to allocate resource");
1822                         GOTO(out, rc);
1823                 }
1824                 LDLM_DEBUG(lock, "completion AST, new resource");
1825                 CERROR("change resource!\n");
1826                 lock_res_and_lock(lock);
1827         }
1828
1829         if (dlm_req->lock_flags & LDLM_FL_AST_SENT) {
1830                 /* BL_AST locks are not needed in LRU.
1831                  * Let ldlm_cancel_lru() be fast. */
1832                 ldlm_lock_remove_from_lru(lock);
1833                 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_BL_AST;
1834                 LDLM_DEBUG(lock, "completion AST includes blocking AST");
1835         }
1836
1837         if (lock->l_lvb_len > 0) {
1838                 rc = ldlm_fill_lvb(lock, &req->rq_pill, RCL_CLIENT,
1839                                    lock->l_lvb_data, lvb_len);
1840                 if (rc < 0) {
1841                         unlock_res_and_lock(lock);
1842                         GOTO(out, rc);
1843                 }
1844         }
1845
1846         ldlm_grant_lock(lock, &ast_list);
1847         unlock_res_and_lock(lock);
1848
1849         LDLM_DEBUG(lock, "callback handler finished, about to run_ast_work");
1850
1851         /* Let Enqueue to call osc_lock_upcall() and initialize
1852          * l_ast_data */
1853         OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_CP_ENQ_RACE, 2);
1854
1855         ldlm_run_ast_work(ns, &ast_list, LDLM_WORK_CP_AST);
1856
1857         LDLM_DEBUG_NOLOCK("client completion callback handler END (lock %p)",
1858                           lock);
1859         GOTO(out, rc);
1860
1861 out:
1862         if (rc < 0) {
1863                 lock_res_and_lock(lock);
1864                 ldlm_set_failed(lock);
1865                 unlock_res_and_lock(lock);
1866                 wake_up(&lock->l_waitq);
1867         }
1868         LDLM_LOCK_RELEASE(lock);
1869 }
1870
1871 /**
1872  * Callback handler for receiving incoming glimpse ASTs.
1873  *
1874  * This only can happen on client side.  After handling the glimpse AST
1875  * we also consider dropping the lock here if it is unused locally for a
1876  * long time.
1877  */
1878 static void ldlm_handle_gl_callback(struct ptlrpc_request *req,
1879                                     struct ldlm_namespace *ns,
1880                                     struct ldlm_request *dlm_req,
1881                                     struct ldlm_lock *lock)
1882 {
1883         int rc = -ENOSYS;
1884         ENTRY;
1885
1886         LDLM_DEBUG(lock, "client glimpse AST callback handler");
1887
1888         if (lock->l_glimpse_ast != NULL)
1889                 rc = lock->l_glimpse_ast(lock, req);
1890
1891         if (req->rq_repmsg != NULL) {
1892                 ptlrpc_reply(req);
1893         } else {
1894                 req->rq_status = rc;
1895                 ptlrpc_error(req);
1896         }
1897
1898         lock_res_and_lock(lock);
1899         if (lock->l_granted_mode == LCK_PW &&
1900             !lock->l_readers && !lock->l_writers &&
1901             cfs_time_after(cfs_time_current(),
1902                            cfs_time_add(lock->l_last_used,
1903                                         cfs_time_seconds(10)))) {
1904                 unlock_res_and_lock(lock);
1905                 if (ldlm_bl_to_thread_lock(ns, NULL, lock))
1906                         ldlm_handle_bl_callback(ns, NULL, lock);
1907
1908                 EXIT;
1909                 return;
1910         }
1911         unlock_res_and_lock(lock);
1912         LDLM_LOCK_RELEASE(lock);
1913         EXIT;
1914 }
1915
1916 static int ldlm_callback_reply(struct ptlrpc_request *req, int rc)
1917 {
1918         if (req->rq_no_reply)
1919                 return 0;
1920
1921         req->rq_status = rc;
1922         if (!req->rq_packed_final) {
1923                 rc = lustre_pack_reply(req, 1, NULL, NULL);
1924                 if (rc)
1925                         return rc;
1926         }
1927         return ptlrpc_reply(req);
1928 }
1929
1930 #ifdef __KERNEL__
1931 static int __ldlm_bl_to_thread(struct ldlm_bl_work_item *blwi,
1932                                ldlm_cancel_flags_t cancel_flags)
1933 {
1934         struct ldlm_bl_pool *blp = ldlm_state->ldlm_bl_pool;
1935         ENTRY;
1936
1937         spin_lock(&blp->blp_lock);
1938         if (blwi->blwi_lock &&
1939             ldlm_is_discard_data(blwi->blwi_lock)) {
1940                 /* add LDLM_FL_DISCARD_DATA requests to the priority list */
1941                 cfs_list_add_tail(&blwi->blwi_entry, &blp->blp_prio_list);
1942         } else {
1943                 /* other blocking callbacks are added to the regular list */
1944                 cfs_list_add_tail(&blwi->blwi_entry, &blp->blp_list);
1945         }
1946         spin_unlock(&blp->blp_lock);
1947
1948         wake_up(&blp->blp_waitq);
1949
1950         /* can not check blwi->blwi_flags as blwi could be already freed in
1951            LCF_ASYNC mode */
1952         if (!(cancel_flags & LCF_ASYNC))
1953                 wait_for_completion(&blwi->blwi_comp);
1954
1955         RETURN(0);
1956 }
1957
1958 static inline void init_blwi(struct ldlm_bl_work_item *blwi,
1959                              struct ldlm_namespace *ns,
1960                              struct ldlm_lock_desc *ld,
1961                              cfs_list_t *cancels, int count,
1962                              struct ldlm_lock *lock,
1963                              ldlm_cancel_flags_t cancel_flags)
1964 {
1965         init_completion(&blwi->blwi_comp);
1966         CFS_INIT_LIST_HEAD(&blwi->blwi_head);
1967
1968         if (memory_pressure_get())
1969                 blwi->blwi_mem_pressure = 1;
1970
1971         blwi->blwi_ns = ns;
1972         blwi->blwi_flags = cancel_flags;
1973         if (ld != NULL)
1974                 blwi->blwi_ld = *ld;
1975         if (count) {
1976                 cfs_list_add(&blwi->blwi_head, cancels);
1977                 cfs_list_del_init(cancels);
1978                 blwi->blwi_count = count;
1979         } else {
1980                 blwi->blwi_lock = lock;
1981         }
1982 }
1983
1984 /**
1985  * Queues a list of locks \a cancels containing \a count locks
1986  * for later processing by a blocking thread.  If \a count is zero,
1987  * then the lock referenced as \a lock is queued instead.
1988  *
1989  * The blocking thread would then call ->l_blocking_ast callback in the lock.
1990  * If list addition fails an error is returned and caller is supposed to
1991  * call ->l_blocking_ast itself.
1992  */
1993 static int ldlm_bl_to_thread(struct ldlm_namespace *ns,
1994                              struct ldlm_lock_desc *ld,
1995                              struct ldlm_lock *lock,
1996                              cfs_list_t *cancels, int count,
1997                              ldlm_cancel_flags_t cancel_flags)
1998 {
1999         ENTRY;
2000
2001         if (cancels && count == 0)
2002                 RETURN(0);
2003
2004         if (cancel_flags & LCF_ASYNC) {
2005                 struct ldlm_bl_work_item *blwi;
2006
2007                 OBD_ALLOC(blwi, sizeof(*blwi));
2008                 if (blwi == NULL)
2009                         RETURN(-ENOMEM);
2010                 init_blwi(blwi, ns, ld, cancels, count, lock, cancel_flags);
2011
2012                 RETURN(__ldlm_bl_to_thread(blwi, cancel_flags));
2013         } else {
2014                 /* if it is synchronous call do minimum mem alloc, as it could
2015                  * be triggered from kernel shrinker
2016                  */
2017                 struct ldlm_bl_work_item blwi;
2018
2019                 memset(&blwi, 0, sizeof(blwi));
2020                 init_blwi(&blwi, ns, ld, cancels, count, lock, cancel_flags);
2021                 RETURN(__ldlm_bl_to_thread(&blwi, cancel_flags));
2022         }
2023 }
2024
2025 #endif
2026
2027 int ldlm_bl_to_thread_lock(struct ldlm_namespace *ns, struct ldlm_lock_desc *ld,
2028                            struct ldlm_lock *lock)
2029 {
2030 #ifdef __KERNEL__
2031         return ldlm_bl_to_thread(ns, ld, lock, NULL, 0, LCF_ASYNC);
2032 #else
2033         return -ENOSYS;
2034 #endif
2035 }
2036
2037 int ldlm_bl_to_thread_list(struct ldlm_namespace *ns, struct ldlm_lock_desc *ld,
2038                            cfs_list_t *cancels, int count,
2039                            ldlm_cancel_flags_t cancel_flags)
2040 {
2041 #ifdef __KERNEL__
2042         return ldlm_bl_to_thread(ns, ld, NULL, cancels, count, cancel_flags);
2043 #else
2044         return -ENOSYS;
2045 #endif
2046 }
2047
2048 /* Setinfo coming from Server (eg MDT) to Client (eg MDC)! */
2049 static int ldlm_handle_setinfo(struct ptlrpc_request *req)
2050 {
2051         struct obd_device *obd = req->rq_export->exp_obd;
2052         char *key;
2053         void *val;
2054         int keylen, vallen;
2055         int rc = -ENOSYS;
2056         ENTRY;
2057
2058         DEBUG_REQ(D_HSM, req, "%s: handle setinfo\n", obd->obd_name);
2059
2060         req_capsule_set(&req->rq_pill, &RQF_OBD_SET_INFO);
2061
2062         key = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_KEY);
2063         if (key == NULL) {
2064                 DEBUG_REQ(D_IOCTL, req, "no set_info key");
2065                 RETURN(-EFAULT);
2066         }
2067         keylen = req_capsule_get_size(&req->rq_pill, &RMF_SETINFO_KEY,
2068                                       RCL_CLIENT);
2069         val = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_VAL);
2070         if (val == NULL) {
2071                 DEBUG_REQ(D_IOCTL, req, "no set_info val");
2072                 RETURN(-EFAULT);
2073         }
2074         vallen = req_capsule_get_size(&req->rq_pill, &RMF_SETINFO_VAL,
2075                                       RCL_CLIENT);
2076
2077         /* We are responsible for swabbing contents of val */
2078
2079         if (KEY_IS(KEY_HSM_COPYTOOL_SEND))
2080                 /* Pass it on to mdc (the "export" in this case) */
2081                 rc = obd_set_info_async(req->rq_svc_thread->t_env,
2082                                         req->rq_export,
2083                                         sizeof(KEY_HSM_COPYTOOL_SEND),
2084                                         KEY_HSM_COPYTOOL_SEND,
2085                                         vallen, val, NULL);
2086         else
2087                 DEBUG_REQ(D_WARNING, req, "ignoring unknown key %s", key);
2088
2089         return rc;
2090 }
2091
2092 static inline void ldlm_callback_errmsg(struct ptlrpc_request *req,
2093                                         const char *msg, int rc,
2094                                         struct lustre_handle *handle)
2095 {
2096         DEBUG_REQ((req->rq_no_reply || rc) ? D_WARNING : D_DLMTRACE, req,
2097                   "%s: [nid %s] [rc %d] [lock "LPX64"]",
2098                   msg, libcfs_id2str(req->rq_peer), rc,
2099                   handle ? handle->cookie : 0);
2100         if (req->rq_no_reply)
2101                 CWARN("No reply was sent, maybe cause bug 21636.\n");
2102         else if (rc)
2103                 CWARN("Send reply failed, maybe cause bug 21636.\n");
2104 }
2105
2106 static int ldlm_handle_qc_callback(struct ptlrpc_request *req)
2107 {
2108         struct obd_quotactl *oqctl;
2109         struct client_obd *cli = &req->rq_export->exp_obd->u.cli;
2110
2111         oqctl = req_capsule_client_get(&req->rq_pill, &RMF_OBD_QUOTACTL);
2112         if (oqctl == NULL) {
2113                 CERROR("Can't unpack obd_quotactl\n");
2114                 RETURN(-EPROTO);
2115         }
2116
2117         oqctl->qc_stat = ptlrpc_status_ntoh(oqctl->qc_stat);
2118
2119         cli->cl_qchk_stat = oqctl->qc_stat;
2120         return 0;
2121 }
2122
2123 /* TODO: handle requests in a similar way as MDT: see mdt_handle_common() */
2124 static int ldlm_callback_handler(struct ptlrpc_request *req)
2125 {
2126         struct ldlm_namespace *ns;
2127         struct ldlm_request *dlm_req;
2128         struct ldlm_lock *lock;
2129         int rc;
2130         ENTRY;
2131
2132         /* Requests arrive in sender's byte order.  The ptlrpc service
2133          * handler has already checked and, if necessary, byte-swapped the
2134          * incoming request message body, but I am responsible for the
2135          * message buffers. */
2136
2137         /* do nothing for sec context finalize */
2138         if (lustre_msg_get_opc(req->rq_reqmsg) == SEC_CTX_FINI)
2139                 RETURN(0);
2140
2141         req_capsule_init(&req->rq_pill, req, RCL_SERVER);
2142
2143         if (req->rq_export == NULL) {
2144                 rc = ldlm_callback_reply(req, -ENOTCONN);
2145                 ldlm_callback_errmsg(req, "Operate on unconnected server",
2146                                      rc, NULL);
2147                 RETURN(0);
2148         }
2149
2150         LASSERT(req->rq_export != NULL);
2151         LASSERT(req->rq_export->exp_obd != NULL);
2152
2153         switch (lustre_msg_get_opc(req->rq_reqmsg)) {
2154         case LDLM_BL_CALLBACK:
2155                 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_BL_CALLBACK_NET))
2156                         RETURN(0);
2157                 break;
2158         case LDLM_CP_CALLBACK:
2159                 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CP_CALLBACK_NET))
2160                         RETURN(0);
2161                 break;
2162         case LDLM_GL_CALLBACK:
2163                 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_GL_CALLBACK_NET))
2164                         RETURN(0);
2165                 break;
2166         case LDLM_SET_INFO:
2167                 rc = ldlm_handle_setinfo(req);
2168                 ldlm_callback_reply(req, rc);
2169                 RETURN(0);
2170         case LLOG_ORIGIN_HANDLE_CREATE:
2171                 req_capsule_set(&req->rq_pill, &RQF_LLOG_ORIGIN_HANDLE_CREATE);
2172                 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOGD_NET))
2173                         RETURN(0);
2174                 rc = llog_origin_handle_open(req);
2175                 ldlm_callback_reply(req, rc);
2176                 RETURN(0);
2177         case LLOG_ORIGIN_HANDLE_NEXT_BLOCK:
2178                 req_capsule_set(&req->rq_pill,
2179                                 &RQF_LLOG_ORIGIN_HANDLE_NEXT_BLOCK);
2180                 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOGD_NET))
2181                         RETURN(0);
2182                 rc = llog_origin_handle_next_block(req);
2183                 ldlm_callback_reply(req, rc);
2184                 RETURN(0);
2185         case LLOG_ORIGIN_HANDLE_READ_HEADER:
2186                 req_capsule_set(&req->rq_pill,
2187                                 &RQF_LLOG_ORIGIN_HANDLE_READ_HEADER);
2188                 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOGD_NET))
2189                         RETURN(0);
2190                 rc = llog_origin_handle_read_header(req);
2191                 ldlm_callback_reply(req, rc);
2192                 RETURN(0);
2193         case LLOG_ORIGIN_HANDLE_CLOSE:
2194                 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOGD_NET))
2195                         RETURN(0);
2196                 rc = llog_origin_handle_close(req);
2197                 ldlm_callback_reply(req, rc);
2198                 RETURN(0);
2199         case OBD_QC_CALLBACK:
2200                 req_capsule_set(&req->rq_pill, &RQF_QC_CALLBACK);
2201                 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_QC_CALLBACK_NET))
2202                         RETURN(0);
2203                 rc = ldlm_handle_qc_callback(req);
2204                 ldlm_callback_reply(req, rc);
2205                 RETURN(0);
2206         default:
2207                 CERROR("unknown opcode %u\n",
2208                        lustre_msg_get_opc(req->rq_reqmsg));
2209                 ldlm_callback_reply(req, -EPROTO);
2210                 RETURN(0);
2211         }
2212
2213         ns = req->rq_export->exp_obd->obd_namespace;
2214         LASSERT(ns != NULL);
2215
2216         req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK);
2217
2218         dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2219         if (dlm_req == NULL) {
2220                 rc = ldlm_callback_reply(req, -EPROTO);
2221                 ldlm_callback_errmsg(req, "Operate without parameter", rc,
2222                                      NULL);
2223                 RETURN(0);
2224         }
2225
2226         /* Force a known safe race, send a cancel to the server for a lock
2227          * which the server has already started a blocking callback on. */
2228         if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_BL_CB_RACE) &&
2229             lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK) {
2230                 rc = ldlm_cli_cancel(&dlm_req->lock_handle[0], 0);
2231                 if (rc < 0)
2232                         CERROR("ldlm_cli_cancel: %d\n", rc);
2233         }
2234
2235         lock = ldlm_handle2lock_long(&dlm_req->lock_handle[0], 0);
2236         if (!lock) {
2237                 CDEBUG(D_DLMTRACE, "callback on lock "LPX64" - lock "
2238                        "disappeared\n", dlm_req->lock_handle[0].cookie);
2239                 rc = ldlm_callback_reply(req, -EINVAL);
2240                 ldlm_callback_errmsg(req, "Operate with invalid parameter", rc,
2241                                      &dlm_req->lock_handle[0]);
2242                 RETURN(0);
2243         }
2244
2245         if (ldlm_is_fail_loc(lock) &&
2246             lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK)
2247                 OBD_RACE(OBD_FAIL_LDLM_CP_BL_RACE);
2248
2249         /* Copy hints/flags (e.g. LDLM_FL_DISCARD_DATA) from AST. */
2250         lock_res_and_lock(lock);
2251         lock->l_flags |= ldlm_flags_from_wire(dlm_req->lock_flags &
2252                                               LDLM_FL_AST_MASK);
2253         if (lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK) {
2254                 /* If somebody cancels lock and cache is already dropped,
2255                  * or lock is failed before cp_ast received on client,
2256                  * we can tell the server we have no lock. Otherwise, we
2257                  * should send cancel after dropping the cache. */
2258                 if ((ldlm_is_canceling(lock) && ldlm_is_bl_done(lock)) ||
2259                     ldlm_is_failed(lock)) {
2260                         LDLM_DEBUG(lock, "callback on lock "
2261                                    LPX64" - lock disappeared\n",
2262                                    dlm_req->lock_handle[0].cookie);
2263                         unlock_res_and_lock(lock);
2264                         LDLM_LOCK_RELEASE(lock);
2265                         rc = ldlm_callback_reply(req, -EINVAL);
2266                         ldlm_callback_errmsg(req, "Operate on stale lock", rc,
2267                                              &dlm_req->lock_handle[0]);
2268                         RETURN(0);
2269                 }
2270                 /* BL_AST locks are not needed in LRU.
2271                  * Let ldlm_cancel_lru() be fast. */
2272                 ldlm_lock_remove_from_lru(lock);
2273                 ldlm_set_bl_ast(lock);
2274         }
2275         unlock_res_and_lock(lock);
2276
2277         /* We want the ost thread to get this reply so that it can respond
2278          * to ost requests (write cache writeback) that might be triggered
2279          * in the callback.
2280          *
2281          * But we'd also like to be able to indicate in the reply that we're
2282          * cancelling right now, because it's unused, or have an intent result
2283          * in the reply, so we might have to push the responsibility for sending
2284          * the reply down into the AST handlers, alas. */
2285
2286         switch (lustre_msg_get_opc(req->rq_reqmsg)) {
2287         case LDLM_BL_CALLBACK:
2288                 CDEBUG(D_INODE, "blocking ast\n");
2289                 req_capsule_extend(&req->rq_pill, &RQF_LDLM_BL_CALLBACK);
2290                 if (!ldlm_is_cancel_on_block(lock)) {
2291                         rc = ldlm_callback_reply(req, 0);
2292                         if (req->rq_no_reply || rc)
2293                                 ldlm_callback_errmsg(req, "Normal process", rc,
2294                                                      &dlm_req->lock_handle[0]);
2295                 }
2296                 if (ldlm_bl_to_thread_lock(ns, &dlm_req->lock_desc, lock))
2297                         ldlm_handle_bl_callback(ns, &dlm_req->lock_desc, lock);
2298                 break;
2299         case LDLM_CP_CALLBACK:
2300                 CDEBUG(D_INODE, "completion ast\n");
2301                 req_capsule_extend(&req->rq_pill, &RQF_LDLM_CP_CALLBACK);
2302                 ldlm_callback_reply(req, 0);
2303                 ldlm_handle_cp_callback(req, ns, dlm_req, lock);
2304                 break;
2305         case LDLM_GL_CALLBACK:
2306                 CDEBUG(D_INODE, "glimpse ast\n");
2307                 req_capsule_extend(&req->rq_pill, &RQF_LDLM_GL_CALLBACK);
2308                 ldlm_handle_gl_callback(req, ns, dlm_req, lock);
2309                 break;
2310         default:
2311                 LBUG();                         /* checked above */
2312         }
2313
2314         RETURN(0);
2315 }
2316
2317 #ifdef HAVE_SERVER_SUPPORT
2318 /**
2319  * Main handler for canceld thread.
2320  *
2321  * Separated into its own thread to avoid deadlocks.
2322  */
2323 static int ldlm_cancel_handler(struct ptlrpc_request *req)
2324 {
2325         int rc;
2326         ENTRY;
2327
2328         /* Requests arrive in sender's byte order.  The ptlrpc service
2329          * handler has already checked and, if necessary, byte-swapped the
2330          * incoming request message body, but I am responsible for the
2331          * message buffers. */
2332
2333         req_capsule_init(&req->rq_pill, req, RCL_SERVER);
2334
2335         if (req->rq_export == NULL) {
2336                 struct ldlm_request *dlm_req;
2337
2338                 CERROR("%s from %s arrived at %lu with bad export cookie "
2339                        LPU64"\n",
2340                        ll_opcode2str(lustre_msg_get_opc(req->rq_reqmsg)),
2341                        libcfs_nid2str(req->rq_peer.nid),
2342                        req->rq_arrival_time.tv_sec,
2343                        lustre_msg_get_handle(req->rq_reqmsg)->cookie);
2344
2345                 if (lustre_msg_get_opc(req->rq_reqmsg) == LDLM_CANCEL) {
2346                         req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK);
2347                         dlm_req = req_capsule_client_get(&req->rq_pill,
2348                                                          &RMF_DLM_REQ);
2349                         if (dlm_req != NULL)
2350                                 ldlm_lock_dump_handle(D_ERROR,
2351                                                       &dlm_req->lock_handle[0]);
2352                 }
2353                 ldlm_callback_reply(req, -ENOTCONN);
2354                 RETURN(0);
2355         }
2356
2357         switch (lustre_msg_get_opc(req->rq_reqmsg)) {
2358
2359         /* XXX FIXME move this back to mds/handler.c, bug 249 */
2360         case LDLM_CANCEL:
2361                 req_capsule_set(&req->rq_pill, &RQF_LDLM_CANCEL);
2362                 CDEBUG(D_INODE, "cancel\n");
2363                 if (CFS_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_NET) ||
2364                     CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_CANCEL_RESEND))
2365                         RETURN(0);
2366                 rc = ldlm_handle_cancel(req);
2367                 if (rc)
2368                         break;
2369                 RETURN(0);
2370         default:
2371                 CERROR("invalid opcode %d\n",
2372                        lustre_msg_get_opc(req->rq_reqmsg));
2373                 req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK);
2374                 ldlm_callback_reply(req, -EINVAL);
2375         }
2376
2377         RETURN(0);
2378 }
2379
2380 static int ldlm_cancel_hpreq_lock_match(struct ptlrpc_request *req,
2381                                         struct ldlm_lock *lock)
2382 {
2383         struct ldlm_request *dlm_req;
2384         struct lustre_handle lockh;
2385         int rc = 0;
2386         int i;
2387         ENTRY;
2388
2389         dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2390         if (dlm_req == NULL)
2391                 RETURN(0);
2392
2393         ldlm_lock2handle(lock, &lockh);
2394         for (i = 0; i < dlm_req->lock_count; i++) {
2395                 if (lustre_handle_equal(&dlm_req->lock_handle[i],
2396                                         &lockh)) {
2397                         DEBUG_REQ(D_RPCTRACE, req,
2398                                   "Prio raised by lock "LPX64".", lockh.cookie);
2399
2400                         rc = 1;
2401                         break;
2402                 }
2403         }
2404
2405         RETURN(rc);
2406
2407 }
2408
2409 static int ldlm_cancel_hpreq_check(struct ptlrpc_request *req)
2410 {
2411         struct ldlm_request *dlm_req;
2412         int rc = 0;
2413         int i;
2414         ENTRY;
2415
2416         /* no prolong in recovery */
2417         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
2418                 RETURN(0);
2419
2420         dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2421         if (dlm_req == NULL)
2422                 RETURN(-EFAULT);
2423
2424         for (i = 0; i < dlm_req->lock_count; i++) {
2425                 struct ldlm_lock *lock;
2426
2427                 lock = ldlm_handle2lock(&dlm_req->lock_handle[i]);
2428                 if (lock == NULL)
2429                         continue;
2430
2431                 rc = ldlm_is_ast_sent(lock) ? 1 : 0;
2432                 if (rc)
2433                         LDLM_DEBUG(lock, "hpreq cancel lock");
2434                 LDLM_LOCK_PUT(lock);
2435
2436                 if (rc)
2437                         break;
2438         }
2439
2440         RETURN(rc);
2441 }
2442
2443 static struct ptlrpc_hpreq_ops ldlm_cancel_hpreq_ops = {
2444         .hpreq_lock_match = ldlm_cancel_hpreq_lock_match,
2445         .hpreq_check      = ldlm_cancel_hpreq_check,
2446         .hpreq_fini       = NULL,
2447 };
2448
2449 static int ldlm_hpreq_handler(struct ptlrpc_request *req)
2450 {
2451         ENTRY;
2452
2453         req_capsule_init(&req->rq_pill, req, RCL_SERVER);
2454
2455         if (req->rq_export == NULL)
2456                 RETURN(0);
2457
2458         if (LDLM_CANCEL == lustre_msg_get_opc(req->rq_reqmsg)) {
2459                 req_capsule_set(&req->rq_pill, &RQF_LDLM_CANCEL);
2460                 req->rq_ops = &ldlm_cancel_hpreq_ops;
2461         }
2462         RETURN(0);
2463 }
2464
2465 int ldlm_revoke_lock_cb(cfs_hash_t *hs, cfs_hash_bd_t *bd,
2466                         cfs_hlist_node_t *hnode, void *data)
2467
2468 {
2469         cfs_list_t         *rpc_list = data;
2470         struct ldlm_lock   *lock = cfs_hash_object(hs, hnode);
2471
2472         lock_res_and_lock(lock);
2473
2474         if (lock->l_req_mode != lock->l_granted_mode) {
2475                 unlock_res_and_lock(lock);
2476                 return 0;
2477         }
2478
2479         LASSERT(lock->l_resource);
2480         if (lock->l_resource->lr_type != LDLM_IBITS &&
2481             lock->l_resource->lr_type != LDLM_PLAIN) {
2482                 unlock_res_and_lock(lock);
2483                 return 0;
2484         }
2485
2486         if (ldlm_is_ast_sent(lock)) {
2487                 unlock_res_and_lock(lock);
2488                 return 0;
2489         }
2490
2491         LASSERT(lock->l_blocking_ast);
2492         LASSERT(!lock->l_blocking_lock);
2493
2494         ldlm_set_ast_sent(lock);
2495         if (lock->l_export && lock->l_export->exp_lock_hash) {
2496                 /* NB: it's safe to call cfs_hash_del() even lock isn't
2497                  * in exp_lock_hash. */
2498                 /* In the function below, .hs_keycmp resolves to
2499                  * ldlm_export_lock_keycmp() */
2500                 /* coverity[overrun-buffer-val] */
2501                 cfs_hash_del(lock->l_export->exp_lock_hash,
2502                              &lock->l_remote_handle, &lock->l_exp_hash);
2503         }
2504
2505         cfs_list_add_tail(&lock->l_rk_ast, rpc_list);
2506         LDLM_LOCK_GET(lock);
2507
2508         unlock_res_and_lock(lock);
2509         return 0;
2510 }
2511
2512 void ldlm_revoke_export_locks(struct obd_export *exp)
2513 {
2514         cfs_list_t  rpc_list;
2515         ENTRY;
2516
2517         CFS_INIT_LIST_HEAD(&rpc_list);
2518         cfs_hash_for_each_empty(exp->exp_lock_hash,
2519                                 ldlm_revoke_lock_cb, &rpc_list);
2520         ldlm_run_ast_work(exp->exp_obd->obd_namespace, &rpc_list,
2521                           LDLM_WORK_REVOKE_AST);
2522
2523         EXIT;
2524 }
2525 EXPORT_SYMBOL(ldlm_revoke_export_locks);
2526 #endif /* HAVE_SERVER_SUPPORT */
2527
2528 #ifdef __KERNEL__
2529 static struct ldlm_bl_work_item *ldlm_bl_get_work(struct ldlm_bl_pool *blp)
2530 {
2531         struct ldlm_bl_work_item *blwi = NULL;
2532         static unsigned int num_bl = 0;
2533
2534         spin_lock(&blp->blp_lock);
2535         /* process a request from the blp_list at least every blp_num_threads */
2536         if (!cfs_list_empty(&blp->blp_list) &&
2537             (cfs_list_empty(&blp->blp_prio_list) || num_bl == 0))
2538                 blwi = cfs_list_entry(blp->blp_list.next,
2539                                       struct ldlm_bl_work_item, blwi_entry);
2540         else
2541                 if (!cfs_list_empty(&blp->blp_prio_list))
2542                         blwi = cfs_list_entry(blp->blp_prio_list.next,
2543                                               struct ldlm_bl_work_item,
2544                                               blwi_entry);
2545
2546         if (blwi) {
2547                 if (++num_bl >= atomic_read(&blp->blp_num_threads))
2548                         num_bl = 0;
2549                 cfs_list_del(&blwi->blwi_entry);
2550         }
2551         spin_unlock(&blp->blp_lock);
2552
2553         return blwi;
2554 }
2555
2556 /* This only contains temporary data until the thread starts */
2557 struct ldlm_bl_thread_data {
2558         char                    bltd_name[CFS_CURPROC_COMM_MAX];
2559         struct ldlm_bl_pool     *bltd_blp;
2560         struct completion       bltd_comp;
2561         int                     bltd_num;
2562 };
2563
2564 static int ldlm_bl_thread_main(void *arg);
2565
2566 static int ldlm_bl_thread_start(struct ldlm_bl_pool *blp)
2567 {
2568         struct ldlm_bl_thread_data bltd = { .bltd_blp = blp };
2569         struct task_struct *task;
2570
2571         init_completion(&bltd.bltd_comp);
2572         bltd.bltd_num = atomic_read(&blp->blp_num_threads);
2573         snprintf(bltd.bltd_name, sizeof(bltd.bltd_name) - 1,
2574                 "ldlm_bl_%02d", bltd.bltd_num);
2575         task = kthread_run(ldlm_bl_thread_main, &bltd, bltd.bltd_name);
2576         if (IS_ERR(task)) {
2577                 CERROR("cannot start LDLM thread ldlm_bl_%02d: rc %ld\n",
2578                        atomic_read(&blp->blp_num_threads), PTR_ERR(task));
2579                 return PTR_ERR(task);
2580         }
2581         wait_for_completion(&bltd.bltd_comp);
2582
2583         return 0;
2584 }
2585
2586 /**
2587  * Main blocking requests processing thread.
2588  *
2589  * Callers put locks into its queue by calling ldlm_bl_to_thread.
2590  * This thread in the end ends up doing actual call to ->l_blocking_ast
2591  * for queued locks.
2592  */
2593 static int ldlm_bl_thread_main(void *arg)
2594 {
2595         struct ldlm_bl_pool *blp;
2596         struct ldlm_bl_thread_data *bltd = arg;
2597         ENTRY;
2598
2599         blp = bltd->bltd_blp;
2600
2601         atomic_inc(&blp->blp_num_threads);
2602         atomic_inc(&blp->blp_busy_threads);
2603
2604         complete(&bltd->bltd_comp);
2605         /* cannot use bltd after this, it is only on caller's stack */
2606
2607         while (1) {
2608                 struct l_wait_info lwi = { 0 };
2609                 struct ldlm_bl_work_item *blwi = NULL;
2610                 int busy;
2611
2612                 blwi = ldlm_bl_get_work(blp);
2613
2614                 if (blwi == NULL) {
2615                         atomic_dec(&blp->blp_busy_threads);
2616                         l_wait_event_exclusive(blp->blp_waitq,
2617                                          (blwi = ldlm_bl_get_work(blp)) != NULL,
2618                                          &lwi);
2619                         busy = atomic_inc_return(&blp->blp_busy_threads);
2620                 } else {
2621                         busy = atomic_read(&blp->blp_busy_threads);
2622                 }
2623
2624                 if (blwi->blwi_ns == NULL)
2625                         /* added by ldlm_cleanup() */
2626                         break;
2627
2628                 /* Not fatal if racy and have a few too many threads */
2629                 if (unlikely(busy < blp->blp_max_threads &&
2630                              busy >= atomic_read(&blp->blp_num_threads) &&
2631                              !blwi->blwi_mem_pressure))
2632                         /* discard the return value, we tried */
2633                         ldlm_bl_thread_start(blp);
2634
2635                 if (blwi->blwi_mem_pressure)
2636                         memory_pressure_set();
2637
2638                 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_PAUSE_CANCEL2, 4);
2639
2640                 if (blwi->blwi_count) {
2641                         int count;
2642                         /* The special case when we cancel locks in LRU
2643                          * asynchronously, we pass the list of locks here.
2644                          * Thus locks are marked LDLM_FL_CANCELING, but NOT
2645                          * canceled locally yet. */
2646                         count = ldlm_cli_cancel_list_local(&blwi->blwi_head,
2647                                                            blwi->blwi_count,
2648                                                            LCF_BL_AST);
2649                         ldlm_cli_cancel_list(&blwi->blwi_head, count, NULL,
2650                                              blwi->blwi_flags);
2651                 } else {
2652                         ldlm_handle_bl_callback(blwi->blwi_ns, &blwi->blwi_ld,
2653                                                 blwi->blwi_lock);
2654                 }
2655                 if (blwi->blwi_mem_pressure)
2656                         memory_pressure_clr();
2657
2658                 if (blwi->blwi_flags & LCF_ASYNC)
2659                         OBD_FREE(blwi, sizeof(*blwi));
2660                 else
2661                         complete(&blwi->blwi_comp);
2662         }
2663
2664         atomic_dec(&blp->blp_busy_threads);
2665         atomic_dec(&blp->blp_num_threads);
2666         complete(&blp->blp_comp);
2667         RETURN(0);
2668 }
2669
2670 #endif
2671
2672 static int ldlm_setup(void);
2673 static int ldlm_cleanup(void);
2674
2675 int ldlm_get_ref(void)
2676 {
2677         int rc = 0;
2678         ENTRY;
2679         mutex_lock(&ldlm_ref_mutex);
2680         if (++ldlm_refcount == 1) {
2681                 rc = ldlm_setup();
2682                 if (rc)
2683                         ldlm_refcount--;
2684         }
2685         mutex_unlock(&ldlm_ref_mutex);
2686
2687         RETURN(rc);
2688 }
2689 EXPORT_SYMBOL(ldlm_get_ref);
2690
2691 void ldlm_put_ref(void)
2692 {
2693         ENTRY;
2694         mutex_lock(&ldlm_ref_mutex);
2695         if (ldlm_refcount == 1) {
2696                 int rc = ldlm_cleanup();
2697                 if (rc)
2698                         CERROR("ldlm_cleanup failed: %d\n", rc);
2699                 else
2700                         ldlm_refcount--;
2701         } else {
2702                 ldlm_refcount--;
2703         }
2704         mutex_unlock(&ldlm_ref_mutex);
2705
2706         EXIT;
2707 }
2708 EXPORT_SYMBOL(ldlm_put_ref);
2709
2710 /*
2711  * Export handle<->lock hash operations.
2712  */
2713 static unsigned
2714 ldlm_export_lock_hash(cfs_hash_t *hs, const void *key, unsigned mask)
2715 {
2716         return cfs_hash_u64_hash(((struct lustre_handle *)key)->cookie, mask);
2717 }
2718
2719 static void *
2720 ldlm_export_lock_key(cfs_hlist_node_t *hnode)
2721 {
2722         struct ldlm_lock *lock;
2723
2724         lock = cfs_hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2725         return &lock->l_remote_handle;
2726 }
2727
2728 static void
2729 ldlm_export_lock_keycpy(cfs_hlist_node_t *hnode, void *key)
2730 {
2731         struct ldlm_lock     *lock;
2732
2733         lock = cfs_hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2734         lock->l_remote_handle = *(struct lustre_handle *)key;
2735 }
2736
2737 static int
2738 ldlm_export_lock_keycmp(const void *key, cfs_hlist_node_t *hnode)
2739 {
2740         return lustre_handle_equal(ldlm_export_lock_key(hnode), key);
2741 }
2742
2743 static void *
2744 ldlm_export_lock_object(cfs_hlist_node_t *hnode)
2745 {
2746         return cfs_hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2747 }
2748
2749 static void
2750 ldlm_export_lock_get(cfs_hash_t *hs, cfs_hlist_node_t *hnode)
2751 {
2752         struct ldlm_lock *lock;
2753
2754         lock = cfs_hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2755         LDLM_LOCK_GET(lock);
2756 }
2757
2758 static void
2759 ldlm_export_lock_put(cfs_hash_t *hs, cfs_hlist_node_t *hnode)
2760 {
2761         struct ldlm_lock *lock;
2762
2763         lock = cfs_hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2764         LDLM_LOCK_RELEASE(lock);
2765 }
2766
2767 static cfs_hash_ops_t ldlm_export_lock_ops = {
2768         .hs_hash        = ldlm_export_lock_hash,
2769         .hs_key         = ldlm_export_lock_key,
2770         .hs_keycmp      = ldlm_export_lock_keycmp,
2771         .hs_keycpy      = ldlm_export_lock_keycpy,
2772         .hs_object      = ldlm_export_lock_object,
2773         .hs_get         = ldlm_export_lock_get,
2774         .hs_put         = ldlm_export_lock_put,
2775         .hs_put_locked  = ldlm_export_lock_put,
2776 };
2777
2778 int ldlm_init_export(struct obd_export *exp)
2779 {
2780         int rc;
2781         ENTRY;
2782
2783         exp->exp_lock_hash =
2784                 cfs_hash_create(obd_uuid2str(&exp->exp_client_uuid),
2785                                 HASH_EXP_LOCK_CUR_BITS,
2786                                 HASH_EXP_LOCK_MAX_BITS,
2787                                 HASH_EXP_LOCK_BKT_BITS, 0,
2788                                 CFS_HASH_MIN_THETA, CFS_HASH_MAX_THETA,
2789                                 &ldlm_export_lock_ops,
2790                                 CFS_HASH_DEFAULT | CFS_HASH_REHASH_KEY |
2791                                 CFS_HASH_NBLK_CHANGE);
2792
2793         if (!exp->exp_lock_hash)
2794                 RETURN(-ENOMEM);
2795
2796         rc = ldlm_init_flock_export(exp);
2797         if (rc)
2798                 GOTO(err, rc);
2799
2800         RETURN(0);
2801 err:
2802         ldlm_destroy_export(exp);
2803         RETURN(rc);
2804 }
2805 EXPORT_SYMBOL(ldlm_init_export);
2806
2807 void ldlm_destroy_export(struct obd_export *exp)
2808 {
2809         ENTRY;
2810         cfs_hash_putref(exp->exp_lock_hash);
2811         exp->exp_lock_hash = NULL;
2812
2813         ldlm_destroy_flock_export(exp);
2814         EXIT;
2815 }
2816 EXPORT_SYMBOL(ldlm_destroy_export);
2817
2818 static int ldlm_setup(void)
2819 {
2820         static struct ptlrpc_service_conf       conf;
2821         struct ldlm_bl_pool                    *blp = NULL;
2822 #ifdef __KERNEL__
2823 # ifdef HAVE_SERVER_SUPPORT
2824         struct task_struct *task;
2825 # endif
2826         int i;
2827 #endif
2828         int rc = 0;
2829
2830         ENTRY;
2831
2832         if (ldlm_state != NULL)
2833                 RETURN(-EALREADY);
2834
2835         OBD_ALLOC(ldlm_state, sizeof(*ldlm_state));
2836         if (ldlm_state == NULL)
2837                 RETURN(-ENOMEM);
2838
2839 #ifdef LPROCFS
2840         rc = ldlm_proc_setup();
2841         if (rc != 0)
2842                 GOTO(out, rc);
2843 #endif
2844
2845         memset(&conf, 0, sizeof(conf));
2846         conf = (typeof(conf)) {
2847                 .psc_name               = "ldlm_cbd",
2848                 .psc_watchdog_factor    = 2,
2849                 .psc_buf                = {
2850                         .bc_nbufs               = LDLM_CLIENT_NBUFS,
2851                         .bc_buf_size            = LDLM_BUFSIZE,
2852                         .bc_req_max_size        = LDLM_MAXREQSIZE,
2853                         .bc_rep_max_size        = LDLM_MAXREPSIZE,
2854                         .bc_req_portal          = LDLM_CB_REQUEST_PORTAL,
2855                         .bc_rep_portal          = LDLM_CB_REPLY_PORTAL,
2856                 },
2857                 .psc_thr                = {
2858                         .tc_thr_name            = "ldlm_cb",
2859                         .tc_thr_factor          = LDLM_THR_FACTOR,
2860                         .tc_nthrs_init          = LDLM_NTHRS_INIT,
2861                         .tc_nthrs_base          = LDLM_NTHRS_BASE,
2862                         .tc_nthrs_max           = LDLM_NTHRS_MAX,
2863                         .tc_nthrs_user          = ldlm_num_threads,
2864                         .tc_cpu_affinity        = 1,
2865                         .tc_ctx_tags            = LCT_MD_THREAD | LCT_DT_THREAD,
2866                 },
2867                 .psc_cpt                = {
2868                         .cc_pattern             = ldlm_cpts,
2869                 },
2870                 .psc_ops                = {
2871                         .so_req_handler         = ldlm_callback_handler,
2872                 },
2873         };
2874         ldlm_state->ldlm_cb_service = \
2875                         ptlrpc_register_service(&conf, ldlm_svc_proc_dir);
2876         if (IS_ERR(ldlm_state->ldlm_cb_service)) {
2877                 CERROR("failed to start service\n");
2878                 rc = PTR_ERR(ldlm_state->ldlm_cb_service);
2879                 ldlm_state->ldlm_cb_service = NULL;
2880                 GOTO(out, rc);
2881         }
2882
2883 #ifdef HAVE_SERVER_SUPPORT
2884         memset(&conf, 0, sizeof(conf));
2885         conf = (typeof(conf)) {
2886                 .psc_name               = "ldlm_canceld",
2887                 .psc_watchdog_factor    = 6,
2888                 .psc_buf                = {
2889                         .bc_nbufs               = LDLM_SERVER_NBUFS,
2890                         .bc_buf_size            = LDLM_BUFSIZE,
2891                         .bc_req_max_size        = LDLM_MAXREQSIZE,
2892                         .bc_rep_max_size        = LDLM_MAXREPSIZE,
2893                         .bc_req_portal          = LDLM_CANCEL_REQUEST_PORTAL,
2894                         .bc_rep_portal          = LDLM_CANCEL_REPLY_PORTAL,
2895
2896                 },
2897                 .psc_thr                = {
2898                         .tc_thr_name            = "ldlm_cn",
2899                         .tc_thr_factor          = LDLM_THR_FACTOR,
2900                         .tc_nthrs_init          = LDLM_NTHRS_INIT,
2901                         .tc_nthrs_base          = LDLM_NTHRS_BASE,
2902                         .tc_nthrs_max           = LDLM_NTHRS_MAX,
2903                         .tc_nthrs_user          = ldlm_num_threads,
2904                         .tc_cpu_affinity        = 1,
2905                         .tc_ctx_tags            = LCT_MD_THREAD | \
2906                                                   LCT_DT_THREAD | \
2907                                                   LCT_CL_THREAD,
2908                 },
2909                 .psc_cpt                = {
2910                         .cc_pattern             = ldlm_cpts,
2911                 },
2912                 .psc_ops                = {
2913                         .so_req_handler         = ldlm_cancel_handler,
2914                         .so_hpreq_handler       = ldlm_hpreq_handler,
2915                 },
2916         };
2917         ldlm_state->ldlm_cancel_service = \
2918                         ptlrpc_register_service(&conf, ldlm_svc_proc_dir);
2919         if (IS_ERR(ldlm_state->ldlm_cancel_service)) {
2920                 CERROR("failed to start service\n");
2921                 rc = PTR_ERR(ldlm_state->ldlm_cancel_service);
2922                 ldlm_state->ldlm_cancel_service = NULL;
2923                 GOTO(out, rc);
2924         }
2925 #endif
2926
2927         OBD_ALLOC(blp, sizeof(*blp));
2928         if (blp == NULL)
2929                 GOTO(out, rc = -ENOMEM);
2930         ldlm_state->ldlm_bl_pool = blp;
2931
2932         spin_lock_init(&blp->blp_lock);
2933         CFS_INIT_LIST_HEAD(&blp->blp_list);
2934         CFS_INIT_LIST_HEAD(&blp->blp_prio_list);
2935         init_waitqueue_head(&blp->blp_waitq);
2936         atomic_set(&blp->blp_num_threads, 0);
2937         atomic_set(&blp->blp_busy_threads, 0);
2938
2939 #ifdef __KERNEL__
2940         if (ldlm_num_threads == 0) {
2941                 blp->blp_min_threads = LDLM_NTHRS_INIT;
2942                 blp->blp_max_threads = LDLM_NTHRS_MAX;
2943         } else {
2944                 blp->blp_min_threads = blp->blp_max_threads = \
2945                         min_t(int, LDLM_NTHRS_MAX, max_t(int, LDLM_NTHRS_INIT,
2946                                                          ldlm_num_threads));
2947         }
2948
2949         for (i = 0; i < blp->blp_min_threads; i++) {
2950                 rc = ldlm_bl_thread_start(blp);
2951                 if (rc < 0)
2952                         GOTO(out, rc);
2953         }
2954
2955 # ifdef HAVE_SERVER_SUPPORT
2956         CFS_INIT_LIST_HEAD(&expired_lock_thread.elt_expired_locks);
2957         expired_lock_thread.elt_state = ELT_STOPPED;
2958         init_waitqueue_head(&expired_lock_thread.elt_waitq);
2959
2960         CFS_INIT_LIST_HEAD(&waiting_locks_list);
2961         spin_lock_init(&waiting_locks_spinlock);
2962         cfs_timer_init(&waiting_locks_timer, waiting_locks_callback, 0);
2963
2964         task = kthread_run(expired_lock_main, NULL, "ldlm_elt");
2965         if (IS_ERR(task)) {
2966                 rc = PTR_ERR(task);
2967                 CERROR("Cannot start ldlm expired-lock thread: %d\n", rc);
2968                 GOTO(out, rc);
2969         }
2970
2971         wait_event(expired_lock_thread.elt_waitq,
2972                        expired_lock_thread.elt_state == ELT_READY);
2973 # endif /* HAVE_SERVER_SUPPORT */
2974
2975         rc = ldlm_pools_init();
2976         if (rc) {
2977                 CERROR("Failed to initialize LDLM pools: %d\n", rc);
2978                 GOTO(out, rc);
2979         }
2980 #endif
2981         RETURN(0);
2982
2983  out:
2984         ldlm_cleanup();
2985         RETURN(rc);
2986 }
2987
2988 static int ldlm_cleanup(void)
2989 {
2990         ENTRY;
2991
2992         if (!cfs_list_empty(ldlm_namespace_list(LDLM_NAMESPACE_SERVER)) ||
2993             !cfs_list_empty(ldlm_namespace_list(LDLM_NAMESPACE_CLIENT))) {
2994                 CERROR("ldlm still has namespaces; clean these up first.\n");
2995                 ldlm_dump_all_namespaces(LDLM_NAMESPACE_SERVER, D_DLMTRACE);
2996                 ldlm_dump_all_namespaces(LDLM_NAMESPACE_CLIENT, D_DLMTRACE);
2997                 RETURN(-EBUSY);
2998         }
2999
3000 #ifdef __KERNEL__
3001         ldlm_pools_fini();
3002
3003         if (ldlm_state->ldlm_bl_pool != NULL) {
3004                 struct ldlm_bl_pool *blp = ldlm_state->ldlm_bl_pool;
3005
3006                 while (atomic_read(&blp->blp_num_threads) > 0) {
3007                         struct ldlm_bl_work_item blwi = { .blwi_ns = NULL };
3008
3009                         init_completion(&blp->blp_comp);
3010
3011                         spin_lock(&blp->blp_lock);
3012                         cfs_list_add_tail(&blwi.blwi_entry, &blp->blp_list);
3013                         wake_up(&blp->blp_waitq);
3014                         spin_unlock(&blp->blp_lock);
3015
3016                         wait_for_completion(&blp->blp_comp);
3017                 }
3018
3019                 OBD_FREE(blp, sizeof(*blp));
3020         }
3021 #endif /* __KERNEL__ */
3022
3023         if (ldlm_state->ldlm_cb_service != NULL)
3024                 ptlrpc_unregister_service(ldlm_state->ldlm_cb_service);
3025 # ifdef HAVE_SERVER_SUPPORT
3026         if (ldlm_state->ldlm_cancel_service != NULL)
3027                 ptlrpc_unregister_service(ldlm_state->ldlm_cancel_service);
3028 # endif
3029
3030 #ifdef __KERNEL__
3031         ldlm_proc_cleanup();
3032
3033 # ifdef HAVE_SERVER_SUPPORT
3034         if (expired_lock_thread.elt_state != ELT_STOPPED) {
3035                 expired_lock_thread.elt_state = ELT_TERMINATE;
3036                 wake_up(&expired_lock_thread.elt_waitq);
3037                 wait_event(expired_lock_thread.elt_waitq,
3038                                expired_lock_thread.elt_state == ELT_STOPPED);
3039         }
3040 # endif
3041 #endif /* __KERNEL__ */
3042
3043         OBD_FREE(ldlm_state, sizeof(*ldlm_state));
3044         ldlm_state = NULL;
3045
3046         RETURN(0);
3047 }
3048
3049 int ldlm_init(void)
3050 {
3051         mutex_init(&ldlm_ref_mutex);
3052         mutex_init(ldlm_namespace_lock(LDLM_NAMESPACE_SERVER));
3053         mutex_init(ldlm_namespace_lock(LDLM_NAMESPACE_CLIENT));
3054         ldlm_resource_slab = kmem_cache_create("ldlm_resources",
3055                                                sizeof(struct ldlm_resource), 0,
3056                                                SLAB_HWCACHE_ALIGN, NULL);
3057         if (ldlm_resource_slab == NULL)
3058                 return -ENOMEM;
3059
3060         ldlm_lock_slab = kmem_cache_create("ldlm_locks",
3061                               sizeof(struct ldlm_lock), 0,
3062                               SLAB_HWCACHE_ALIGN | SLAB_DESTROY_BY_RCU, NULL);
3063         if (ldlm_lock_slab == NULL) {
3064                 kmem_cache_destroy(ldlm_resource_slab);
3065                 return -ENOMEM;
3066         }
3067
3068         ldlm_interval_slab = kmem_cache_create("interval_node",
3069                                         sizeof(struct ldlm_interval),
3070                                         0, SLAB_HWCACHE_ALIGN, NULL);
3071         if (ldlm_interval_slab == NULL) {
3072                 kmem_cache_destroy(ldlm_resource_slab);
3073                 kmem_cache_destroy(ldlm_lock_slab);
3074                 return -ENOMEM;
3075         }
3076 #if LUSTRE_TRACKS_LOCK_EXP_REFS
3077         class_export_dump_hook = ldlm_dump_export_locks;
3078 #endif
3079         return 0;
3080 }
3081
3082 void ldlm_exit(void)
3083 {
3084         if (ldlm_refcount)
3085                 CERROR("ldlm_refcount is %d in ldlm_exit!\n", ldlm_refcount);
3086         kmem_cache_destroy(ldlm_resource_slab);
3087 #ifdef __KERNEL__
3088         /* ldlm_lock_put() use RCU to call ldlm_lock_free, so need call
3089          * synchronize_rcu() to wait a grace period elapsed, so that
3090          * ldlm_lock_free() get a chance to be called. */
3091         synchronize_rcu();
3092 #endif
3093         kmem_cache_destroy(ldlm_lock_slab);
3094         kmem_cache_destroy(ldlm_interval_slab);
3095 }