Whamcloud - gitweb
LU-12931 ldlm: use proper units for timeouts
[fs/lustre-release.git] / lustre / ldlm / ldlm_lockd.c
1 /*
2  * GPL HEADER START
3  *
4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
5  *
6  * This program is free software; you can redistribute it and/or modify
7  * it under the terms of the GNU General Public License version 2 only,
8  * as published by the Free Software Foundation.
9  *
10  * This program is distributed in the hope that it will be useful, but
11  * WITHOUT ANY WARRANTY; without even the implied warranty of
12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
13  * General Public License version 2 for more details (a copy is included
14  * in the LICENSE file that accompanied this code).
15  *
16  * You should have received a copy of the GNU General Public License
17  * version 2 along with this program; If not, see
18  * http://www.gnu.org/licenses/gpl-2.0.html
19  *
20  * GPL HEADER END
21  */
22 /*
23  * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
24  * Use is subject to license terms.
25  *
26  * Copyright (c) 2010, 2017, Intel Corporation.
27  */
28 /*
29  * This file is part of Lustre, http://www.lustre.org/
30  * Lustre is a trademark of Sun Microsystems, Inc.
31  *
32  * lustre/ldlm/ldlm_lockd.c
33  *
34  * Author: Peter Braam <braam@clusterfs.com>
35  * Author: Phil Schwan <phil@clusterfs.com>
36  */
37
38 #define DEBUG_SUBSYSTEM S_LDLM
39
40 #include <linux/kthread.h>
41 #include <linux/list.h>
42 #include <libcfs/libcfs.h>
43 #include <libcfs/linux/linux-mem.h>
44 #include <lustre_errno.h>
45 #include <lustre_dlm.h>
46 #include <obd_class.h>
47 #include "ldlm_internal.h"
48
49 static int ldlm_num_threads;
50 module_param(ldlm_num_threads, int, 0444);
51 MODULE_PARM_DESC(ldlm_num_threads, "number of DLM service threads to start");
52
53 static unsigned int ldlm_cpu_bind = 1;
54 module_param(ldlm_cpu_bind, uint, 0444);
55 MODULE_PARM_DESC(ldlm_cpu_bind,
56                  "bind DLM service threads to particular CPU partitions");
57
58 static char *ldlm_cpts;
59 module_param(ldlm_cpts, charp, 0444);
60 MODULE_PARM_DESC(ldlm_cpts, "CPU partitions ldlm threads should run on");
61
62 static DEFINE_MUTEX(ldlm_ref_mutex);
63 static int ldlm_refcount;
64
65 struct kobject *ldlm_kobj;
66 struct kset *ldlm_ns_kset;
67 struct kset *ldlm_svc_kset;
68
69 /* LDLM state */
70
71 static struct ldlm_state *ldlm_state;
72
73 /*
74  * timeout for initial callback (AST) reply (bz10399)
75  * Due to having to send a 32 bit time value over the
76  * wire return it as timeout_t instead of time64_t
77  */
78 static inline timeout_t ldlm_get_rq_timeout(void)
79 {
80         /* Non-AT value */
81         timeout_t timeout = min(ldlm_timeout, obd_timeout / 3);
82
83         return timeout < 1 ? 1 : timeout;
84 }
85
86 struct ldlm_bl_pool {
87         spinlock_t blp_lock;
88
89         /*
90          * blp_prio_list is used for callbacks that should be handled
91          * as a priority. It is used for LDLM_FL_DISCARD_DATA requests.
92          * see b=13843
93          */
94         struct list_head blp_prio_list;
95
96         /*
97          * blp_list is used for all other callbacks which are likely
98          * to take longer to process.
99          */
100         struct list_head blp_list;
101
102         wait_queue_head_t blp_waitq;
103         struct completion blp_comp;
104         atomic_t blp_num_threads;
105         atomic_t blp_busy_threads;
106         int blp_min_threads;
107         int blp_max_threads;
108 };
109
110 struct ldlm_bl_work_item {
111         struct list_head        blwi_entry;
112         struct ldlm_namespace   *blwi_ns;
113         struct ldlm_lock_desc   blwi_ld;
114         struct ldlm_lock        *blwi_lock;
115         struct list_head        blwi_head;
116         int                     blwi_count;
117         struct completion       blwi_comp;
118         enum ldlm_cancel_flags  blwi_flags;
119         int                     blwi_mem_pressure;
120 };
121
122 #ifdef HAVE_SERVER_SUPPORT
123
124 /**
125  * Protects both waiting_locks_list and expired_lock_thread.
126  */
127 static DEFINE_SPINLOCK(waiting_locks_spinlock); /* BH lock (timer) */
128
129 /**
130  * List for contended locks.
131  *
132  * As soon as a lock is contended, it gets placed on this list and
133  * expected time to get a response is filled in the lock. A special
134  * thread walks the list looking for locks that should be released and
135  * schedules client evictions for those that have not been released in
136  * time.
137  *
138  * All access to it should be under waiting_locks_spinlock.
139  */
140 static LIST_HEAD(waiting_locks_list);
141 static void waiting_locks_callback(TIMER_DATA_TYPE unused);
142 static CFS_DEFINE_TIMER(waiting_locks_timer, waiting_locks_callback, 0, 0);
143
144 enum elt_state {
145         ELT_STOPPED,
146         ELT_READY,
147         ELT_TERMINATE,
148 };
149
150 static DECLARE_WAIT_QUEUE_HEAD(expired_lock_wait_queue);
151 static enum elt_state expired_lock_thread_state = ELT_STOPPED;
152 static int expired_lock_dump;
153 static LIST_HEAD(expired_lock_list);
154
155 static int ldlm_lock_busy(struct ldlm_lock *lock);
156 static int ldlm_add_waiting_lock(struct ldlm_lock *lock, timeout_t timeout);
157 static int __ldlm_add_waiting_lock(struct ldlm_lock *lock, timeout_t timeout);
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 = !list_empty(&expired_lock_list);
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         struct list_head *expired = &expired_lock_list;
177         int do_dump;
178
179         ENTRY;
180
181         expired_lock_thread_state = ELT_READY;
182         wake_up(&expired_lock_wait_queue);
183
184         while (1) {
185                 wait_event_idle(expired_lock_wait_queue,
186                                 have_expired_locks() ||
187                                 expired_lock_thread_state == ELT_TERMINATE);
188
189                 spin_lock_bh(&waiting_locks_spinlock);
190                 if (expired_lock_dump) {
191                         spin_unlock_bh(&waiting_locks_spinlock);
192
193                         /* from waiting_locks_callback, but not in timer */
194                         libcfs_debug_dumplog();
195
196                         spin_lock_bh(&waiting_locks_spinlock);
197                         expired_lock_dump = 0;
198                 }
199
200                 do_dump = 0;
201
202                 while (!list_empty(expired)) {
203                         struct obd_export *export;
204                         struct ldlm_lock *lock;
205
206                         lock = list_entry(expired->next, struct ldlm_lock,
207                                           l_pending_chain);
208                         if ((void *)lock < LP_POISON + PAGE_SIZE &&
209                             (void *)lock >= LP_POISON) {
210                                 spin_unlock_bh(&waiting_locks_spinlock);
211                                 CERROR("free lock on elt list %p\n", lock);
212                                 LBUG();
213                         }
214                         list_del_init(&lock->l_pending_chain);
215                         if ((void *)lock->l_export <
216                              LP_POISON + PAGE_SIZE &&
217                             (void *)lock->l_export >= LP_POISON) {
218                                 CERROR("lock with free export on elt list %p\n",
219                                        lock->l_export);
220                                 lock->l_export = NULL;
221                                 LDLM_ERROR(lock, "free export");
222                                 /*
223                                  * release extra ref grabbed by
224                                  * ldlm_add_waiting_lock() or
225                                  * ldlm_failed_ast()
226                                  */
227                                 LDLM_LOCK_RELEASE(lock);
228                                 continue;
229                         }
230
231                         if (ldlm_is_destroyed(lock)) {
232                                 /*
233                                  * release the lock refcount where
234                                  * waiting_locks_callback() founds
235                                  */
236                                 LDLM_LOCK_RELEASE(lock);
237                                 continue;
238                         }
239                         export = class_export_lock_get(lock->l_export, lock);
240                         spin_unlock_bh(&waiting_locks_spinlock);
241
242                         /* Check if we need to prolong timeout */
243                         if (!OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_TIMEOUT) &&
244                             lock->l_callback_timestamp != 0 && /* not AST error */
245                             ldlm_lock_busy(lock)) {
246                                 LDLM_DEBUG(lock, "prolong the busy lock");
247                                 lock_res_and_lock(lock);
248                                 ldlm_add_waiting_lock(lock,
249                                                 ldlm_bl_timeout(lock) >> 1);
250                                 unlock_res_and_lock(lock);
251                         } else {
252                                 spin_lock_bh(&export->exp_bl_list_lock);
253                                 list_del_init(&lock->l_exp_list);
254                                 spin_unlock_bh(&export->exp_bl_list_lock);
255
256                                 LDLM_ERROR(lock,
257                                            "lock callback timer expired after %llds: evicting client at %s ",
258                                            ktime_get_real_seconds() -
259                                            lock->l_blast_sent,
260                                            obd_export_nid2str(export));
261                                 ldlm_lock_to_ns(lock)->ns_timeouts++;
262                                 do_dump++;
263                                 class_fail_export(export);
264                         }
265                         class_export_lock_put(export, lock);
266                         /*
267                          * release extra ref grabbed by ldlm_add_waiting_lock()
268                          * or ldlm_failed_ast()
269                          */
270                         LDLM_LOCK_RELEASE(lock);
271
272                         spin_lock_bh(&waiting_locks_spinlock);
273                 }
274                 spin_unlock_bh(&waiting_locks_spinlock);
275
276                 if (do_dump && obd_dump_on_eviction) {
277                         CERROR("dump the log upon eviction\n");
278                         libcfs_debug_dumplog();
279                 }
280
281                 if (expired_lock_thread_state == ELT_TERMINATE)
282                         break;
283         }
284
285         expired_lock_thread_state = ELT_STOPPED;
286         wake_up(&expired_lock_wait_queue);
287         RETURN(0);
288 }
289
290 /**
291  * Check if there is a request in the export request list
292  * which prevents the lock canceling.
293  */
294 static int ldlm_lock_busy(struct ldlm_lock *lock)
295 {
296         struct ptlrpc_request *req;
297         int match = 0;
298
299         ENTRY;
300
301         if (lock->l_export == NULL)
302                 return 0;
303
304         spin_lock(&lock->l_export->exp_rpc_lock);
305         list_for_each_entry(req, &lock->l_export->exp_hp_rpcs,
306                                 rq_exp_list) {
307                 if (req->rq_ops->hpreq_lock_match) {
308                         match = req->rq_ops->hpreq_lock_match(req, lock);
309                         if (match)
310                                 break;
311                 }
312         }
313         spin_unlock(&lock->l_export->exp_rpc_lock);
314         RETURN(match);
315 }
316
317 /* This is called from within a timer interrupt and cannot schedule */
318 static void waiting_locks_callback(TIMER_DATA_TYPE unused)
319 {
320         struct ldlm_lock *lock;
321         int need_dump = 0;
322
323         spin_lock_bh(&waiting_locks_spinlock);
324         while (!list_empty(&waiting_locks_list)) {
325                 lock = list_entry(waiting_locks_list.next, struct ldlm_lock,
326                                   l_pending_chain);
327                 if (lock->l_callback_timestamp > ktime_get_seconds() ||
328                     lock->l_req_mode == LCK_GROUP)
329                         break;
330
331                 /*
332                  * no needs to take an extra ref on the lock since it was in
333                  * the waiting_locks_list and ldlm_add_waiting_lock()
334                  * already grabbed a ref
335                  */
336                 list_move(&lock->l_pending_chain, &expired_lock_list);
337                 need_dump = 1;
338         }
339
340         if (!list_empty(&expired_lock_list)) {
341                 if (obd_dump_on_timeout && need_dump)
342                         expired_lock_dump = __LINE__;
343
344                 wake_up(&expired_lock_wait_queue);
345         }
346
347         /*
348          * Make sure the timer will fire again if we have any locks
349          * left.
350          */
351         if (!list_empty(&waiting_locks_list)) {
352                 time64_t now = ktime_get_seconds();
353                 timeout_t delta = 0;
354
355                 lock = list_entry(waiting_locks_list.next, struct ldlm_lock,
356                                   l_pending_chain);
357                 if (lock->l_callback_timestamp - now > 0)
358                         delta = lock->l_callback_timestamp - now;
359                 mod_timer(&waiting_locks_timer,
360                           jiffies + cfs_time_seconds(delta));
361         }
362         spin_unlock_bh(&waiting_locks_spinlock);
363 }
364
365 /**
366  * Add lock to the list of contended locks.
367  *
368  * Indicate that we're waiting for a client to call us back cancelling a given
369  * lock.  We add it to the pending-callback chain, and schedule the lock-timeout
370  * timer to fire appropriately.  (We round up to the next second, to avoid
371  * floods of timer firings during periods of high lock contention and traffic).
372  * As done by ldlm_add_waiting_lock(), the caller must grab a lock reference
373  * if it has been added to the waiting list (1 is returned).
374  *
375  * Called with the namespace lock held.
376  */
377 static int __ldlm_add_waiting_lock(struct ldlm_lock *lock, timeout_t delay)
378 {
379         unsigned long timeout_jiffies = jiffies;
380         time64_t now = ktime_get_seconds();
381         time64_t deadline;
382         timeout_t timeout;
383
384         if (!list_empty(&lock->l_pending_chain))
385                 return 0;
386
387         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_NOTIMEOUT) ||
388             OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_TIMEOUT))
389                 delay = 1;
390
391         deadline = now + delay;
392         if (likely(deadline > lock->l_callback_timestamp))
393                 lock->l_callback_timestamp = deadline;
394
395         timeout = clamp_t(timeout_t, lock->l_callback_timestamp - now,
396                           0, delay);
397         timeout_jiffies += cfs_time_seconds(timeout);
398
399         if (time_before(timeout_jiffies, waiting_locks_timer.expires) ||
400             !timer_pending(&waiting_locks_timer))
401                 mod_timer(&waiting_locks_timer, timeout_jiffies);
402
403         /*
404          * if the new lock has a shorter timeout than something earlier on
405          * the list, we'll wait the longer amount of time; no big deal.
406          */
407         /* FIFO */
408         list_add_tail(&lock->l_pending_chain, &waiting_locks_list);
409         return 1;
410 }
411
412 static void ldlm_add_blocked_lock(struct ldlm_lock *lock)
413 {
414         spin_lock_bh(&lock->l_export->exp_bl_list_lock);
415         if (list_empty(&lock->l_exp_list)) {
416                 if (!ldlm_is_granted(lock))
417                         list_add_tail(&lock->l_exp_list,
418                                       &lock->l_export->exp_bl_list);
419                 else
420                         list_add(&lock->l_exp_list,
421                                  &lock->l_export->exp_bl_list);
422         }
423         spin_unlock_bh(&lock->l_export->exp_bl_list_lock);
424
425         /*
426          * A blocked lock is added. Adjust the position in
427          * the stale list if the export is in the list.
428          * If export is stale and not in the list - it is being
429          * processed and will be placed on the right position
430          * on obd_stale_export_put().
431          */
432         if (!list_empty(&lock->l_export->exp_stale_list))
433                 obd_stale_export_adjust(lock->l_export);
434 }
435
436 static int ldlm_add_waiting_lock(struct ldlm_lock *lock, timeout_t timeout)
437 {
438         int ret;
439
440         /* NB: must be called with hold of lock_res_and_lock() */
441         LASSERT(ldlm_is_res_locked(lock));
442         LASSERT(!ldlm_is_cancel_on_block(lock));
443
444         /*
445          * Do not put cross-MDT lock in the waiting list, since we
446          * will not evict it due to timeout for now
447          */
448         if (lock->l_export != NULL &&
449             (exp_connect_flags(lock->l_export) & OBD_CONNECT_MDS_MDS))
450                 return 0;
451
452         spin_lock_bh(&waiting_locks_spinlock);
453         if (ldlm_is_cancel(lock)) {
454                 spin_unlock_bh(&waiting_locks_spinlock);
455                 return 0;
456         }
457
458         if (ldlm_is_destroyed(lock)) {
459                 static time64_t next;
460
461                 spin_unlock_bh(&waiting_locks_spinlock);
462                 LDLM_ERROR(lock, "not waiting on destroyed lock (b=5653)");
463                 if (ktime_get_seconds() > next) {
464                         next = ktime_get_seconds() + 14400;
465                         libcfs_debug_dumpstack(NULL);
466                 }
467                 return 0;
468         }
469
470         ldlm_set_waited(lock);
471         lock->l_blast_sent = ktime_get_real_seconds();
472         ret = __ldlm_add_waiting_lock(lock, timeout);
473         if (ret) {
474                 /*
475                  * grab ref on the lock if it has been added to the
476                  * waiting list
477                  */
478                 LDLM_LOCK_GET(lock);
479         }
480         spin_unlock_bh(&waiting_locks_spinlock);
481
482         if (ret)
483                 ldlm_add_blocked_lock(lock);
484
485         LDLM_DEBUG(lock, "%sadding to wait list(timeout: %d, AT: %s)",
486                    ret == 0 ? "not re-" : "", timeout,
487                    AT_OFF ? "off" : "on");
488         return ret;
489 }
490
491 /**
492  * Remove a lock from the pending list, likely because it had its cancellation
493  * callback arrive without incident.  This adjusts the lock-timeout timer if
494  * needed.  Returns 0 if the lock wasn't pending after all, 1 if it was.
495  * As done by ldlm_del_waiting_lock(), the caller must release the lock
496  * reference when the lock is removed from any list (1 is returned).
497  *
498  * Called with namespace lock held.
499  */
500 static int __ldlm_del_waiting_lock(struct ldlm_lock *lock)
501 {
502         struct list_head *list_next;
503
504         if (list_empty(&lock->l_pending_chain))
505                 return 0;
506
507         list_next = lock->l_pending_chain.next;
508         if (lock->l_pending_chain.prev == &waiting_locks_list) {
509                 /* Removing the head of the list, adjust timer. */
510                 if (list_next == &waiting_locks_list) {
511                         /* No more, just cancel. */
512                         del_timer(&waiting_locks_timer);
513                 } else {
514                         time64_t now = ktime_get_seconds();
515                         struct ldlm_lock *next;
516                         timeout_t delta = 0;
517
518                         next = list_entry(list_next, struct ldlm_lock,
519                                           l_pending_chain);
520                         if (next->l_callback_timestamp - now > 0)
521                                 delta = lock->l_callback_timestamp - now;
522
523                         mod_timer(&waiting_locks_timer,
524                                   jiffies + cfs_time_seconds(delta));
525                 }
526         }
527         list_del_init(&lock->l_pending_chain);
528
529         return 1;
530 }
531
532 int ldlm_del_waiting_lock(struct ldlm_lock *lock)
533 {
534         int ret;
535
536         if (lock->l_export == NULL) {
537                 /* We don't have a "waiting locks list" on clients. */
538                 CDEBUG(D_DLMTRACE, "Client lock %p : no-op\n", lock);
539                 return 0;
540         }
541
542         spin_lock_bh(&waiting_locks_spinlock);
543         ret = __ldlm_del_waiting_lock(lock);
544         ldlm_clear_waited(lock);
545         spin_unlock_bh(&waiting_locks_spinlock);
546
547         /* remove the lock out of export blocking list */
548         spin_lock_bh(&lock->l_export->exp_bl_list_lock);
549         list_del_init(&lock->l_exp_list);
550         spin_unlock_bh(&lock->l_export->exp_bl_list_lock);
551
552         if (ret) {
553                 /*
554                  * release lock ref if it has indeed been removed
555                  * from a list
556                  */
557                 LDLM_LOCK_RELEASE(lock);
558         }
559
560         LDLM_DEBUG(lock, "%s", ret == 0 ? "wasn't waiting" : "removed");
561         return ret;
562 }
563
564 /**
565  * Prolong the contended lock waiting time.
566  *
567  * Called with namespace lock held.
568  */
569 int ldlm_refresh_waiting_lock(struct ldlm_lock *lock, timeout_t timeout)
570 {
571         if (lock->l_export == NULL) {
572                 /* We don't have a "waiting locks list" on clients. */
573                 LDLM_DEBUG(lock, "client lock: no-op");
574                 return 0;
575         }
576
577         if (exp_connect_flags(lock->l_export) & OBD_CONNECT_MDS_MDS) {
578                 /* We don't have a "waiting locks list" on OSP. */
579                 LDLM_DEBUG(lock, "MDS-MDS lock: no-op");
580                 return 0;
581         }
582
583         spin_lock_bh(&waiting_locks_spinlock);
584
585         if (list_empty(&lock->l_pending_chain)) {
586                 spin_unlock_bh(&waiting_locks_spinlock);
587                 LDLM_DEBUG(lock, "wasn't waiting");
588                 return 0;
589         }
590
591         /*
592          * we remove/add the lock to the waiting list, so no needs to
593          * release/take a lock reference
594          */
595         __ldlm_del_waiting_lock(lock);
596         __ldlm_add_waiting_lock(lock, timeout);
597         spin_unlock_bh(&waiting_locks_spinlock);
598
599         LDLM_DEBUG(lock, "refreshed");
600         return 1;
601 }
602 EXPORT_SYMBOL(ldlm_refresh_waiting_lock);
603
604 #else /* HAVE_SERVER_SUPPORT */
605
606 int ldlm_del_waiting_lock(struct ldlm_lock *lock)
607 {
608         RETURN(0);
609 }
610
611 int ldlm_refresh_waiting_lock(struct ldlm_lock *lock, timeout_t timeout)
612 {
613         RETURN(0);
614 }
615
616 #endif /* !HAVE_SERVER_SUPPORT */
617
618 #ifdef HAVE_SERVER_SUPPORT
619
620 /**
621  * Calculate the per-export Blocking timeout (covering BL AST, data flush,
622  * lock cancel, and their replies). Used for lock callback timeout and AST
623  * re-send period.
624  *
625  * \param[in] lock        lock which is getting the blocking callback
626  *
627  * \retval            timeout in seconds to wait for the client reply
628  */
629 timeout_t ldlm_bl_timeout(struct ldlm_lock *lock)
630 {
631         timeout_t timeout;
632
633         if (AT_OFF)
634                 return obd_timeout / 2;
635
636         /*
637          * Since these are non-updating timeouts, we should be conservative.
638          * Take more than usually, 150%
639          * It would be nice to have some kind of "early reply" mechanism for
640          * lock callbacks too...
641          */
642         timeout = at_get(&lock->l_export->exp_bl_lock_at);
643         return max_t(timeout_t, timeout + (timeout >> 1),
644                      (timeout_t)ldlm_enqueue_min);
645 }
646 EXPORT_SYMBOL(ldlm_bl_timeout);
647
648 /**
649  * Perform lock cleanup if AST sending failed.
650  */
651 static void ldlm_failed_ast(struct ldlm_lock *lock, int rc,
652                             const char *ast_type)
653 {
654         LCONSOLE_ERROR_MSG(0x138,
655                            "%s: A client on nid %s was evicted due to a lock %s callback time out: rc %d\n",
656                            lock->l_export->exp_obd->obd_name,
657                            obd_export_nid2str(lock->l_export), ast_type, rc);
658
659         if (obd_dump_on_timeout)
660                 libcfs_debug_dumplog();
661         spin_lock_bh(&waiting_locks_spinlock);
662         if (__ldlm_del_waiting_lock(lock) == 0)
663                 /*
664                  * the lock was not in any list, grab an extra ref before adding
665                  * the lock to the expired list
666                  */
667                 LDLM_LOCK_GET(lock);
668         /* differentiate it from expired locks */
669         lock->l_callback_timestamp = 0;
670         list_add(&lock->l_pending_chain, &expired_lock_list);
671         wake_up(&expired_lock_wait_queue);
672         spin_unlock_bh(&waiting_locks_spinlock);
673 }
674
675 /**
676  * Perform lock cleanup if AST reply came with error.
677  */
678 static int ldlm_handle_ast_error(struct ldlm_lock *lock,
679                                  struct ptlrpc_request *req, int rc,
680                                  const char *ast_type)
681 {
682         struct lnet_process_id peer = req->rq_import->imp_connection->c_peer;
683
684         if (!req->rq_replied || (rc && rc != -EINVAL)) {
685                 if (ldlm_is_cancel(lock)) {
686                         LDLM_DEBUG(lock,
687                                    "%s AST (req@%p x%llu) timeout from nid %s, but cancel was received (AST reply lost?)",
688                                    ast_type, req, req->rq_xid,
689                                    libcfs_nid2str(peer.nid));
690                         ldlm_lock_cancel(lock);
691                         rc = -ERESTART;
692                 } else if (rc == -ENODEV || rc == -ESHUTDOWN ||
693                            (rc == -EIO &&
694                             req->rq_import->imp_state == LUSTRE_IMP_CLOSED)) {
695                         /*
696                          * Upon umount process the AST fails because cannot be
697                          * sent. This shouldn't lead to the client eviction.
698                          * -ENODEV error is returned by ptl_send_rpc() for
699                          *  new request in such import.
700                          * -SHUTDOWN is returned by ptlrpc_import_delay_req()
701                          *  if imp_invalid is set or obd_no_recov.
702                          * Meanwhile there is also check for LUSTRE_IMP_CLOSED
703                          * in ptlrpc_import_delay_req() as well with -EIO code.
704                          * In all such cases errors are ignored.
705                          */
706                         LDLM_DEBUG(lock,
707                                    "%s AST can't be sent due to a server %s failure or umount process: rc = %d\n",
708                                     ast_type,
709                                      req->rq_import->imp_obd->obd_name, rc);
710                 } else {
711                         LDLM_ERROR(lock,
712                                    "client (nid %s) %s %s AST (req@%p x%llu status %d rc %d), evict it",
713                                    libcfs_nid2str(peer.nid),
714                                    req->rq_replied ? "returned error from" :
715                                    "failed to reply to",
716                                    ast_type, req, req->rq_xid,
717                                    (req->rq_repmsg != NULL) ?
718                                    lustre_msg_get_status(req->rq_repmsg) : 0,
719                                    rc);
720                         ldlm_failed_ast(lock, rc, ast_type);
721                 }
722                 return rc;
723         }
724
725         if (rc == -EINVAL) {
726                 struct ldlm_resource *res = lock->l_resource;
727
728                 LDLM_DEBUG(lock,
729                            "client (nid %s) returned %d from %s AST (req@%p x%llu) - normal race",
730                            libcfs_nid2str(peer.nid),
731                            req->rq_repmsg ?
732                            lustre_msg_get_status(req->rq_repmsg) : -1,
733                            ast_type, req, req->rq_xid);
734                 if (res) {
735                         /*
736                          * update lvbo to return proper attributes.
737                          * see b=23174
738                          */
739                         ldlm_resource_getref(res);
740                         ldlm_lvbo_update(res, lock, NULL, 1);
741                         ldlm_resource_putref(res);
742                 }
743                 ldlm_lock_cancel(lock);
744                 rc = -ERESTART;
745         }
746
747         return rc;
748 }
749
750 static int ldlm_cb_interpret(const struct lu_env *env,
751                              struct ptlrpc_request *req, void *args, int rc)
752 {
753         struct ldlm_cb_async_args *ca = args;
754         struct ldlm_lock *lock = ca->ca_lock;
755         struct ldlm_cb_set_arg *arg  = ca->ca_set_arg;
756
757         ENTRY;
758
759         LASSERT(lock != NULL);
760
761         switch (arg->type) {
762         case LDLM_GL_CALLBACK:
763                 /*
764                  * Update the LVB from disk if the AST failed
765                  * (this is a legal race)
766                  *
767                  * - Glimpse callback of local lock just returns
768                  *   -ELDLM_NO_LOCK_DATA.
769                  * - Glimpse callback of remote lock might return
770                  *   -ELDLM_NO_LOCK_DATA when inode is cleared. LU-274
771                  */
772                 if (unlikely(arg->gl_interpret_reply)) {
773                         rc = arg->gl_interpret_reply(NULL, req, args, rc);
774                 } else if (rc == -ELDLM_NO_LOCK_DATA) {
775                         LDLM_DEBUG(lock,
776                                    "lost race - client has a lock but no inode");
777                         ldlm_lvbo_update(lock->l_resource, lock, NULL, 1);
778                 } else if (rc != 0) {
779                         rc = ldlm_handle_ast_error(lock, req, rc, "glimpse");
780                 } else {
781                         rc = ldlm_lvbo_update(lock->l_resource,
782                                               lock, req, 1);
783                 }
784                 break;
785         case LDLM_BL_CALLBACK:
786                 if (rc != 0)
787                         rc = ldlm_handle_ast_error(lock, req, rc, "blocking");
788                 break;
789         case LDLM_CP_CALLBACK:
790                 if (rc != 0)
791                         rc = ldlm_handle_ast_error(lock, req, rc, "completion");
792                 break;
793         default:
794                 LDLM_ERROR(lock, "invalid opcode for lock callback %d",
795                            arg->type);
796                 LBUG();
797         }
798
799         /* release extra reference taken in ldlm_ast_fini() */
800         LDLM_LOCK_RELEASE(lock);
801
802         if (rc == -ERESTART)
803                 atomic_inc(&arg->restart);
804
805         RETURN(0);
806 }
807
808 static void ldlm_update_resend(struct ptlrpc_request *req, void *data)
809 {
810         struct ldlm_cb_async_args *ca = data;
811         struct ldlm_lock *lock = ca->ca_lock;
812
813         ldlm_refresh_waiting_lock(lock, ldlm_bl_timeout(lock));
814 }
815
816 static inline int ldlm_ast_fini(struct ptlrpc_request *req,
817                                 struct ldlm_cb_set_arg *arg,
818                                 struct ldlm_lock *lock,
819                                 int instant_cancel)
820 {
821         int rc = 0;
822
823         ENTRY;
824
825         if (unlikely(instant_cancel)) {
826                 rc = ptl_send_rpc(req, 1);
827                 ptlrpc_req_finished(req);
828                 if (rc == 0)
829                         atomic_inc(&arg->restart);
830         } else {
831                 LDLM_LOCK_GET(lock);
832                 ptlrpc_set_add_req(arg->set, req);
833         }
834
835         RETURN(rc);
836 }
837
838 /**
839  * Check if there are requests in the export request list which prevent
840  * the lock canceling and make these requests high priority ones.
841  */
842 static void ldlm_lock_reorder_req(struct ldlm_lock *lock)
843 {
844         struct ptlrpc_request *req;
845
846         ENTRY;
847
848         if (lock->l_export == NULL) {
849                 LDLM_DEBUG(lock, "client lock: no-op");
850                 RETURN_EXIT;
851         }
852
853         spin_lock(&lock->l_export->exp_rpc_lock);
854         list_for_each_entry(req, &lock->l_export->exp_hp_rpcs,
855                             rq_exp_list) {
856                 /*
857                  * Do not process requests that were not yet added to there
858                  * incoming queue or were already removed from there for
859                  * processing. We evaluate ptlrpc_nrs_req_can_move() without
860                  * holding svcpt->scp_req_lock, and then redo the check with
861                  * the lock held once we need to obtain a reliable result.
862                  */
863                 if (ptlrpc_nrs_req_can_move(req) &&
864                     req->rq_ops->hpreq_lock_match &&
865                     req->rq_ops->hpreq_lock_match(req, lock))
866                         ptlrpc_nrs_req_hp_move(req);
867         }
868         spin_unlock(&lock->l_export->exp_rpc_lock);
869         EXIT;
870 }
871
872 /**
873  * ->l_blocking_ast() method for server-side locks. This is invoked when newly
874  * enqueued server lock conflicts with given one.
875  *
876  * Sends blocking AST RPC to the client owning that lock; arms timeout timer
877  * to wait for client response.
878  */
879 int ldlm_server_blocking_ast(struct ldlm_lock *lock,
880                              struct ldlm_lock_desc *desc,
881                              void *data, int flag)
882 {
883         struct ldlm_cb_async_args *ca;
884         struct ldlm_cb_set_arg *arg = data;
885         struct ldlm_request *body;
886         struct ptlrpc_request  *req;
887         int instant_cancel = 0;
888         int rc = 0;
889
890         ENTRY;
891
892         if (flag == LDLM_CB_CANCELING)
893                 /* Don't need to do anything here. */
894                 RETURN(0);
895
896         if (OBD_FAIL_PRECHECK(OBD_FAIL_LDLM_SRV_BL_AST)) {
897                 LDLM_DEBUG(lock, "dropping BL AST");
898                 RETURN(0);
899         }
900
901         LASSERT(lock);
902         LASSERT(data != NULL);
903         if (lock->l_export->exp_obd->obd_recovering != 0)
904                 LDLM_ERROR(lock, "BUG 6063: lock collide during recovery");
905
906         ldlm_lock_reorder_req(lock);
907
908         req = ptlrpc_request_alloc_pack(lock->l_export->exp_imp_reverse,
909                                         &RQF_LDLM_BL_CALLBACK,
910                                         LUSTRE_DLM_VERSION, LDLM_BL_CALLBACK);
911         if (req == NULL)
912                 RETURN(-ENOMEM);
913
914         ca = ptlrpc_req_async_args(ca, req);
915         ca->ca_set_arg = arg;
916         ca->ca_lock = lock;
917
918         req->rq_interpret_reply = ldlm_cb_interpret;
919
920         lock_res_and_lock(lock);
921         if (ldlm_is_destroyed(lock)) {
922                 /* What's the point? */
923                 unlock_res_and_lock(lock);
924                 ptlrpc_req_finished(req);
925                 RETURN(0);
926         }
927
928         if (!ldlm_is_granted(lock)) {
929                 /*
930                  * this blocking AST will be communicated as part of the
931                  * completion AST instead
932                  */
933                 ldlm_add_blocked_lock(lock);
934                 ldlm_set_waited(lock);
935                 unlock_res_and_lock(lock);
936
937                 ptlrpc_req_finished(req);
938                 LDLM_DEBUG(lock, "lock not granted, not sending blocking AST");
939                 RETURN(0);
940         }
941
942         if (ldlm_is_cancel_on_block(lock))
943                 instant_cancel = 1;
944
945         body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
946         body->lock_handle[0] = lock->l_remote_handle;
947         body->lock_desc = *desc;
948         body->lock_flags |= ldlm_flags_to_wire(lock->l_flags & LDLM_FL_AST_MASK);
949
950         LDLM_DEBUG(lock, "server preparing blocking AST");
951
952         ptlrpc_request_set_replen(req);
953         ldlm_set_cbpending(lock);
954         if (instant_cancel) {
955                 unlock_res_and_lock(lock);
956                 ldlm_lock_cancel(lock);
957
958                 req->rq_no_resend = 1;
959         } else {
960                 LASSERT(ldlm_is_granted(lock));
961                 ldlm_add_waiting_lock(lock, ldlm_bl_timeout(lock));
962                 unlock_res_and_lock(lock);
963
964                 /* Do not resend after lock callback timeout */
965                 req->rq_delay_limit = ldlm_bl_timeout(lock);
966                 req->rq_resend_cb = ldlm_update_resend;
967         }
968
969         req->rq_send_state = LUSTRE_IMP_FULL;
970         /* ptlrpc_request_alloc_pack already set timeout */
971         if (AT_OFF)
972                 req->rq_timeout = ldlm_get_rq_timeout();
973
974         if (lock->l_export && lock->l_export->exp_nid_stats &&
975             lock->l_export->exp_nid_stats->nid_ldlm_stats)
976                 lprocfs_counter_incr(lock->l_export->exp_nid_stats->nid_ldlm_stats,
977                                      LDLM_BL_CALLBACK - LDLM_FIRST_OPC);
978
979         rc = ldlm_ast_fini(req, arg, lock, instant_cancel);
980
981         RETURN(rc);
982 }
983
984 /**
985  * ->l_completion_ast callback for a remote lock in server namespace.
986  *
987  *  Sends AST to the client notifying it of lock granting.  If initial
988  *  lock response was not sent yet, instead of sending another RPC, just
989  *  mark the lock as granted and client will understand
990  */
991 int ldlm_server_completion_ast(struct ldlm_lock *lock, __u64 flags, void *data)
992 {
993         struct ldlm_cb_set_arg *arg = data;
994         struct ldlm_request *body;
995         struct ptlrpc_request *req;
996         struct ldlm_cb_async_args *ca;
997         int instant_cancel = 0;
998         int rc = 0;
999         int lvb_len;
1000
1001         ENTRY;
1002
1003         LASSERT(lock != NULL);
1004         LASSERT(data != NULL);
1005
1006         if (OBD_FAIL_PRECHECK(OBD_FAIL_LDLM_SRV_CP_AST)) {
1007                 LDLM_DEBUG(lock, "dropping CP AST");
1008                 RETURN(0);
1009         }
1010
1011         req = ptlrpc_request_alloc(lock->l_export->exp_imp_reverse,
1012                                    &RQF_LDLM_CP_CALLBACK);
1013         if (req == NULL)
1014                 RETURN(-ENOMEM);
1015
1016         /* server namespace, doesn't need lock */
1017         lvb_len = ldlm_lvbo_size(lock);
1018         /*
1019          * LU-3124 & LU-2187: to not return layout in completion AST because
1020          * it may deadlock for LU-2187, or client may not have enough space
1021          * for large layout. The layout will be returned to client with an
1022          * extra RPC to fetch xattr.lov
1023          */
1024         if (ldlm_has_layout(lock))
1025                 lvb_len = 0;
1026
1027         req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_CLIENT, lvb_len);
1028         rc = ptlrpc_request_pack(req, LUSTRE_DLM_VERSION, LDLM_CP_CALLBACK);
1029         if (rc) {
1030                 ptlrpc_request_free(req);
1031                 RETURN(rc);
1032         }
1033
1034         ca = ptlrpc_req_async_args(ca, req);
1035         ca->ca_set_arg = arg;
1036         ca->ca_lock = lock;
1037
1038         req->rq_interpret_reply = ldlm_cb_interpret;
1039         body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1040
1041         body->lock_handle[0] = lock->l_remote_handle;
1042         body->lock_flags = ldlm_flags_to_wire(flags);
1043         ldlm_lock2desc(lock, &body->lock_desc);
1044         if (lvb_len > 0) {
1045                 void *lvb = req_capsule_client_get(&req->rq_pill, &RMF_DLM_LVB);
1046                 lvb_len = ldlm_lvbo_fill(lock, lvb, &lvb_len);
1047                 if (lvb_len < 0) {
1048                         /*
1049                          * We still need to send the RPC to wake up the blocked
1050                          * enqueue thread on the client.
1051                          *
1052                          * Consider old client, there is no better way to notify
1053                          * the failure, just zero-sized the LVB, then the client
1054                          * will fail out as "-EPROTO".
1055                          */
1056                         req_capsule_shrink(&req->rq_pill, &RMF_DLM_LVB, 0,
1057                                            RCL_CLIENT);
1058                         instant_cancel = 1;
1059                 } else {
1060                         req_capsule_shrink(&req->rq_pill, &RMF_DLM_LVB, lvb_len,
1061                                            RCL_CLIENT);
1062                 }
1063         }
1064
1065         LDLM_DEBUG(lock, "server preparing completion AST");
1066
1067         ptlrpc_request_set_replen(req);
1068
1069         req->rq_send_state = LUSTRE_IMP_FULL;
1070         /* ptlrpc_request_pack already set timeout */
1071         if (AT_OFF)
1072                 req->rq_timeout = ldlm_get_rq_timeout();
1073
1074         /* We only send real blocking ASTs after the lock is granted */
1075         lock_res_and_lock(lock);
1076         if (ldlm_is_ast_sent(lock)) {
1077                 body->lock_flags |= ldlm_flags_to_wire(LDLM_FL_AST_SENT);
1078                 /* Copy AST flags like LDLM_FL_DISCARD_DATA. */
1079                 body->lock_flags |= ldlm_flags_to_wire(lock->l_flags &
1080                                                        LDLM_FL_AST_MASK);
1081
1082                 /*
1083                  * We might get here prior to ldlm_handle_enqueue setting
1084                  * LDLM_FL_CANCEL_ON_BLOCK flag. Then we will put this lock
1085                  * into waiting list, but this is safe and similar code in
1086                  * ldlm_handle_enqueue will call ldlm_lock_cancel() still,
1087                  * that would not only cancel the lock, but will also remove
1088                  * it from waiting list
1089                  */
1090                 if (ldlm_is_cancel_on_block(lock)) {
1091                         unlock_res_and_lock(lock);
1092                         ldlm_lock_cancel(lock);
1093
1094                         instant_cancel = 1;
1095                         req->rq_no_resend = 1;
1096
1097                         lock_res_and_lock(lock);
1098                 } else {
1099                         /* start the lock-timeout clock */
1100                         ldlm_add_waiting_lock(lock, ldlm_bl_timeout(lock));
1101                         /* Do not resend after lock callback timeout */
1102                         req->rq_delay_limit = ldlm_bl_timeout(lock);
1103                         req->rq_resend_cb = ldlm_update_resend;
1104                 }
1105         }
1106         unlock_res_and_lock(lock);
1107
1108         if (lock->l_export && lock->l_export->exp_nid_stats &&
1109             lock->l_export->exp_nid_stats->nid_ldlm_stats)
1110                 lprocfs_counter_incr(lock->l_export->exp_nid_stats->nid_ldlm_stats,
1111                                      LDLM_CP_CALLBACK - LDLM_FIRST_OPC);
1112
1113         rc = ldlm_ast_fini(req, arg, lock, instant_cancel);
1114
1115         RETURN(lvb_len < 0 ? lvb_len : rc);
1116 }
1117
1118 /**
1119  * Server side ->l_glimpse_ast handler for client locks.
1120  *
1121  * Sends glimpse AST to the client and waits for reply. Then updates
1122  * lvbo with the result.
1123  */
1124 int ldlm_server_glimpse_ast(struct ldlm_lock *lock, void *data)
1125 {
1126         struct ldlm_cb_set_arg *arg = data;
1127         struct ldlm_request *body;
1128         struct ptlrpc_request *req;
1129         struct ldlm_cb_async_args *ca;
1130         int rc;
1131         struct req_format *req_fmt;
1132
1133         ENTRY;
1134
1135         LASSERT(lock != NULL);
1136
1137         if (arg->gl_desc != NULL)
1138                 /* There is a glimpse descriptor to pack */
1139                 req_fmt = &RQF_LDLM_GL_CALLBACK_DESC;
1140         else
1141                 req_fmt = &RQF_LDLM_GL_CALLBACK;
1142
1143         req = ptlrpc_request_alloc_pack(lock->l_export->exp_imp_reverse,
1144                                         req_fmt, LUSTRE_DLM_VERSION,
1145                                         LDLM_GL_CALLBACK);
1146
1147         if (req == NULL)
1148                 RETURN(-ENOMEM);
1149
1150         if (arg->gl_desc != NULL) {
1151                 /* copy the GL descriptor */
1152                 union ldlm_gl_desc      *desc;
1153
1154                 desc = req_capsule_client_get(&req->rq_pill, &RMF_DLM_GL_DESC);
1155                 *desc = *arg->gl_desc;
1156         }
1157
1158         body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1159         body->lock_handle[0] = lock->l_remote_handle;
1160         ldlm_lock2desc(lock, &body->lock_desc);
1161
1162         ca = ptlrpc_req_async_args(ca, req);
1163         ca->ca_set_arg = arg;
1164         ca->ca_lock = lock;
1165
1166         /* server namespace, doesn't need lock */
1167         req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
1168                              ldlm_lvbo_size(lock));
1169         ptlrpc_request_set_replen(req);
1170
1171         req->rq_send_state = LUSTRE_IMP_FULL;
1172         /* ptlrpc_request_alloc_pack already set timeout */
1173         if (AT_OFF)
1174                 req->rq_timeout = ldlm_get_rq_timeout();
1175
1176         req->rq_interpret_reply = ldlm_cb_interpret;
1177
1178         if (lock->l_export && lock->l_export->exp_nid_stats) {
1179                 struct nid_stat *nid_stats = lock->l_export->exp_nid_stats;
1180
1181                 lprocfs_counter_incr(nid_stats->nid_ldlm_stats,
1182                                      LDLM_GL_CALLBACK - LDLM_FIRST_OPC);
1183         }
1184
1185         rc = ldlm_ast_fini(req, arg, lock, 0);
1186
1187         RETURN(rc);
1188 }
1189 EXPORT_SYMBOL(ldlm_server_glimpse_ast);
1190
1191 int ldlm_glimpse_locks(struct ldlm_resource *res,
1192                        struct list_head *gl_work_list)
1193 {
1194         int rc;
1195
1196         ENTRY;
1197
1198         rc = ldlm_run_ast_work(ldlm_res_to_ns(res), gl_work_list,
1199                                LDLM_WORK_GL_AST);
1200         if (rc == -ERESTART)
1201                 ldlm_reprocess_all(res, NULL);
1202
1203         RETURN(rc);
1204 }
1205 EXPORT_SYMBOL(ldlm_glimpse_locks);
1206
1207 /* return LDLM lock associated with a lock callback request */
1208 struct ldlm_lock *ldlm_request_lock(struct ptlrpc_request *req)
1209 {
1210         struct ldlm_cb_async_args *ca;
1211         struct ldlm_lock *lock;
1212
1213         ENTRY;
1214
1215         ca = ptlrpc_req_async_args(ca, req);
1216         lock = ca->ca_lock;
1217         if (lock == NULL)
1218                 RETURN(ERR_PTR(-EFAULT));
1219
1220         RETURN(lock);
1221 }
1222 EXPORT_SYMBOL(ldlm_request_lock);
1223
1224 /**
1225  * Main server-side entry point into LDLM for enqueue. This is called by ptlrpc
1226  * service threads to carry out client lock enqueueing requests.
1227  */
1228 int ldlm_handle_enqueue0(struct ldlm_namespace *ns,
1229                          struct ptlrpc_request *req,
1230                          const struct ldlm_request *dlm_req,
1231                          const struct ldlm_callback_suite *cbs)
1232 {
1233         struct ldlm_reply *dlm_rep;
1234         __u64 flags;
1235         enum ldlm_error err = ELDLM_OK;
1236         struct ldlm_lock *lock = NULL;
1237         void *cookie = NULL;
1238         int rc = 0;
1239         struct ldlm_resource *res = NULL;
1240         const struct lu_env *env = req->rq_svc_thread->t_env;
1241
1242         ENTRY;
1243
1244         LDLM_DEBUG_NOLOCK("server-side enqueue handler START");
1245
1246         ldlm_request_cancel(req, dlm_req, LDLM_ENQUEUE_CANCEL_OFF, LATF_SKIP);
1247         flags = ldlm_flags_from_wire(dlm_req->lock_flags);
1248
1249         LASSERT(req->rq_export);
1250
1251         /* for intent enqueue the stat will be updated inside intent policy */
1252         if (ptlrpc_req2svc(req)->srv_stats != NULL &&
1253             !(dlm_req->lock_flags & LDLM_FL_HAS_INTENT))
1254                 ldlm_svc_get_eopc(dlm_req, ptlrpc_req2svc(req)->srv_stats);
1255
1256         if (req->rq_export && req->rq_export->exp_nid_stats &&
1257             req->rq_export->exp_nid_stats->nid_ldlm_stats)
1258                 lprocfs_counter_incr(req->rq_export->exp_nid_stats->nid_ldlm_stats,
1259                                      LDLM_ENQUEUE - LDLM_FIRST_OPC);
1260
1261         if (unlikely(dlm_req->lock_desc.l_resource.lr_type < LDLM_MIN_TYPE ||
1262                      dlm_req->lock_desc.l_resource.lr_type >= LDLM_MAX_TYPE)) {
1263                 DEBUG_REQ(D_ERROR, req, "invalid lock request type %d",
1264                           dlm_req->lock_desc.l_resource.lr_type);
1265                 GOTO(out, rc = -EFAULT);
1266         }
1267
1268         if (unlikely(dlm_req->lock_desc.l_req_mode <= LCK_MINMODE ||
1269                      dlm_req->lock_desc.l_req_mode >= LCK_MAXMODE ||
1270                      dlm_req->lock_desc.l_req_mode &
1271                      (dlm_req->lock_desc.l_req_mode-1))) {
1272                 DEBUG_REQ(D_ERROR, req, "invalid lock request mode %d",
1273                           dlm_req->lock_desc.l_req_mode);
1274                 GOTO(out, rc = -EFAULT);
1275         }
1276
1277         if (unlikely((flags & LDLM_FL_REPLAY) ||
1278                      (lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT))) {
1279                 /* Find an existing lock in the per-export lock hash */
1280                 /*
1281                  * In the function below, .hs_keycmp resolves to
1282                  * ldlm_export_lock_keycmp()
1283                  */
1284                 /* coverity[overrun-buffer-val] */
1285                 lock = cfs_hash_lookup(req->rq_export->exp_lock_hash,
1286                                        (void *)&dlm_req->lock_handle[0]);
1287                 if (lock != NULL) {
1288                         DEBUG_REQ(D_DLMTRACE, req,
1289                                   "found existing lock cookie %#llx",
1290                                   lock->l_handle.h_cookie);
1291                         flags |= LDLM_FL_RESENT;
1292                         GOTO(existing_lock, rc = 0);
1293                 }
1294         } else {
1295                 if (ldlm_reclaim_full()) {
1296                         DEBUG_REQ(D_DLMTRACE, req,
1297                                   "Too many granted locks, reject current enqueue request and let the client retry later");
1298                         GOTO(out, rc = -EINPROGRESS);
1299                 }
1300         }
1301
1302         /* The lock's callback data might be set in the policy function */
1303         lock = ldlm_lock_create(ns, &dlm_req->lock_desc.l_resource.lr_name,
1304                                 dlm_req->lock_desc.l_resource.lr_type,
1305                                 dlm_req->lock_desc.l_req_mode,
1306                                 cbs, NULL, 0, LVB_T_NONE);
1307         if (IS_ERR(lock)) {
1308                 rc = PTR_ERR(lock);
1309                 lock = NULL;
1310                 GOTO(out, rc);
1311         }
1312
1313         lock->l_remote_handle = dlm_req->lock_handle[0];
1314         LDLM_DEBUG(lock, "server-side enqueue handler, new lock created");
1315
1316         /*
1317          * Initialize resource lvb but not for a lock being replayed since
1318          * Client already got lvb sent in this case.
1319          * This must occur early since some policy methods assume resource
1320          * lvb is available (lr_lvb_data != NULL).
1321          */
1322         res = lock->l_resource;
1323         if (!(flags & LDLM_FL_REPLAY)) {
1324                 /* non-replayed lock, delayed lvb init may need to be done */
1325                 rc = ldlm_lvbo_init(res);
1326                 if (rc < 0) {
1327                         LDLM_DEBUG(lock, "delayed lvb init failed (rc %d)", rc);
1328                         GOTO(out, rc);
1329                 }
1330         }
1331
1332         OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_ENQUEUE_BLOCKED, obd_timeout * 2);
1333         /*
1334          * Don't enqueue a lock onto the export if it is been disonnected
1335          * due to eviction (b=3822) or server umount (b=24324).
1336          * Cancel it now instead.
1337          */
1338         if (req->rq_export->exp_disconnected) {
1339                 LDLM_ERROR(lock, "lock on disconnected export %p",
1340                            req->rq_export);
1341                 GOTO(out, rc = -ENOTCONN);
1342         }
1343
1344         lock->l_export = class_export_lock_get(req->rq_export, lock);
1345         if (lock->l_export->exp_lock_hash)
1346                 cfs_hash_add(lock->l_export->exp_lock_hash,
1347                              &lock->l_remote_handle,
1348                              &lock->l_exp_hash);
1349
1350         /*
1351          * Inherit the enqueue flags before the operation, because we do not
1352          * keep the res lock on return and next operations (BL AST) may proceed
1353          * without them.
1354          */
1355         lock->l_flags |= ldlm_flags_from_wire(dlm_req->lock_flags &
1356                                               LDLM_FL_INHERIT_MASK);
1357
1358         ldlm_convert_policy_to_local(req->rq_export,
1359                                      dlm_req->lock_desc.l_resource.lr_type,
1360                                      &dlm_req->lock_desc.l_policy_data,
1361                                      &lock->l_policy_data);
1362         if (dlm_req->lock_desc.l_resource.lr_type == LDLM_EXTENT)
1363                 lock->l_req_extent = lock->l_policy_data.l_extent;
1364         else if (dlm_req->lock_desc.l_resource.lr_type == LDLM_IBITS)
1365                 lock->l_policy_data.l_inodebits.try_bits =
1366                         dlm_req->lock_desc.l_policy_data.l_inodebits.try_bits;
1367
1368 existing_lock:
1369         cookie = req;
1370         if (!(flags & LDLM_FL_HAS_INTENT)) {
1371                 /* based on the assumption that lvb size never changes during
1372                  * resource life time otherwise it need resource->lr_lock's
1373                  * protection */
1374                 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB,
1375                                      RCL_SERVER, ldlm_lvbo_size(lock));
1376
1377                 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_ENQUEUE_EXTENT_ERR))
1378                         GOTO(out, rc = -ENOMEM);
1379
1380                 rc = req_capsule_server_pack(&req->rq_pill);
1381                 if (rc)
1382                         GOTO(out, rc);
1383         }
1384
1385         err = ldlm_lock_enqueue(env, ns, &lock, cookie, &flags);
1386         if (err) {
1387                 if ((int)err < 0)
1388                         rc = (int)err;
1389                 GOTO(out, err);
1390         }
1391
1392         dlm_rep = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
1393
1394         ldlm_lock2desc(lock, &dlm_rep->lock_desc);
1395         ldlm_lock2handle(lock, &dlm_rep->lock_handle);
1396
1397         if (lock && lock->l_resource->lr_type == LDLM_EXTENT)
1398                 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_BL_EVICT, 6);
1399
1400         /*
1401          * We never send a blocking AST until the lock is granted, but
1402          * we can tell it right now
1403          */
1404         lock_res_and_lock(lock);
1405
1406         /*
1407          * Now take into account flags to be inherited from original lock
1408          * request both in reply to client and in our own lock flags.
1409          */
1410         dlm_rep->lock_flags = ldlm_flags_to_wire(flags);
1411         lock->l_flags |= flags & LDLM_FL_INHERIT_MASK;
1412
1413         /*
1414          * Don't move a pending lock onto the export if it has already been
1415          * disconnected due to eviction (b=5683) or server umount (b=24324).
1416          * Cancel it now instead.
1417          */
1418         if (unlikely(req->rq_export->exp_disconnected ||
1419                      OBD_FAIL_CHECK(OBD_FAIL_LDLM_ENQUEUE_OLD_EXPORT))) {
1420                 LDLM_ERROR(lock, "lock on destroyed export %p", req->rq_export);
1421                 rc = -ENOTCONN;
1422         } else if (ldlm_is_ast_sent(lock)) {
1423                 /* fill lock desc for possible lock convert */
1424                 if (lock->l_blocking_lock &&
1425                     lock->l_resource->lr_type == LDLM_IBITS) {
1426                         struct ldlm_lock *bl_lock = lock->l_blocking_lock;
1427                         struct ldlm_lock_desc *rep_desc = &dlm_rep->lock_desc;
1428
1429                         LDLM_DEBUG(lock,
1430                                    "save blocking bits %llx in granted lock",
1431                                    bl_lock->l_policy_data.l_inodebits.bits);
1432                         /*
1433                          * If lock is blocked then save blocking ibits
1434                          * in returned lock policy for the possible lock
1435                          * convert on a client.
1436                          */
1437                         rep_desc->l_policy_data.l_inodebits.cancel_bits =
1438                                 bl_lock->l_policy_data.l_inodebits.bits;
1439                 }
1440                 dlm_rep->lock_flags |= ldlm_flags_to_wire(LDLM_FL_AST_SENT);
1441                 if (ldlm_is_granted(lock)) {
1442                         /*
1443                          * Only cancel lock if it was granted, because it would
1444                          * be destroyed immediately and would never be granted
1445                          * in the future, causing timeouts on client.  Not
1446                          * granted lock will be cancelled immediately after
1447                          * sending completion AST.
1448                          */
1449                         if (ldlm_is_cancel_on_block(lock)) {
1450                                 unlock_res_and_lock(lock);
1451                                 ldlm_lock_cancel(lock);
1452                                 lock_res_and_lock(lock);
1453                         } else {
1454                                 ldlm_add_waiting_lock(lock,
1455                                                       ldlm_bl_timeout(lock));
1456                         }
1457                 }
1458         }
1459         unlock_res_and_lock(lock);
1460
1461         EXIT;
1462 out:
1463         req->rq_status = rc ?: err; /* return either error - b=11190 */
1464         if (!req->rq_packed_final) {
1465                 err = lustre_pack_reply(req, 1, NULL, NULL);
1466                 if (rc == 0)
1467                         rc = err;
1468         }
1469
1470         /*
1471          * The LOCK_CHANGED code in ldlm_lock_enqueue depends on this
1472          * ldlm_reprocess_all.  If this moves, revisit that code. -phil
1473          */
1474         if (lock != NULL) {
1475                 LDLM_DEBUG(lock,
1476                            "server-side enqueue handler, sending reply (err=%d, rc=%d)",
1477                            err, rc);
1478
1479                 if (rc == 0 &&
1480                     req_capsule_has_field(&req->rq_pill, &RMF_DLM_LVB,
1481                                           RCL_SERVER) &&
1482                     ldlm_lvbo_size(lock) > 0) {
1483                         void *buf;
1484                         int buflen;
1485
1486 retry:
1487                         buf = req_capsule_server_get(&req->rq_pill,
1488                                                      &RMF_DLM_LVB);
1489                         LASSERTF(buf != NULL, "req %p, lock %p\n", req, lock);
1490                         buflen = req_capsule_get_size(&req->rq_pill,
1491                                         &RMF_DLM_LVB, RCL_SERVER);
1492                         /*
1493                          * non-replayed lock, delayed lvb init may
1494                          * need to be occur now
1495                          */
1496                         if ((buflen > 0) && !(flags & LDLM_FL_REPLAY)) {
1497                                 int rc2;
1498
1499                                 rc2 = ldlm_lvbo_fill(lock, buf, &buflen);
1500                                 if (rc2 >= 0) {
1501                                         req_capsule_shrink(&req->rq_pill,
1502                                                            &RMF_DLM_LVB,
1503                                                            rc2, RCL_SERVER);
1504                                 } else if (rc2 == -ERANGE) {
1505                                         rc2 = req_capsule_server_grow(
1506                                                         &req->rq_pill,
1507                                                         &RMF_DLM_LVB, buflen);
1508                                         if (!rc2) {
1509                                                 goto retry;
1510                                         } else {
1511                                                 /*
1512                                                  * if we can't grow the buffer,
1513                                                  * it's ok to return empty lvb
1514                                                  * to client.
1515                                                  */
1516                                                 req_capsule_shrink(
1517                                                         &req->rq_pill,
1518                                                         &RMF_DLM_LVB, 0,
1519                                                         RCL_SERVER);
1520                                         }
1521                                 } else {
1522                                         rc = rc2;
1523                                 }
1524                         } else if (flags & LDLM_FL_REPLAY) {
1525                                 /* no LVB resend upon replay */
1526                                 if (buflen > 0)
1527                                         req_capsule_shrink(&req->rq_pill,
1528                                                            &RMF_DLM_LVB,
1529                                                            0, RCL_SERVER);
1530                                 else
1531                                         rc = buflen;
1532                         } else {
1533                                 rc = buflen;
1534                         }
1535                 }
1536
1537                 if (rc != 0 && !(flags & LDLM_FL_RESENT)) {
1538                         if (lock->l_export) {
1539                                 ldlm_lock_cancel(lock);
1540                         } else {
1541                                 lock_res_and_lock(lock);
1542                                 ldlm_resource_unlink_lock(lock);
1543                                 ldlm_lock_destroy_nolock(lock);
1544                                 unlock_res_and_lock(lock);
1545
1546                         }
1547                 }
1548
1549                 if (!err && !ldlm_is_cbpending(lock) &&
1550                     dlm_req->lock_desc.l_resource.lr_type != LDLM_FLOCK)
1551                         ldlm_reprocess_all(lock->l_resource, lock);
1552
1553                 LDLM_LOCK_RELEASE(lock);
1554         }
1555
1556         LDLM_DEBUG_NOLOCK("server-side enqueue handler END (lock %p, rc %d)",
1557                           lock, rc);
1558
1559         return rc;
1560 }
1561
1562 /*
1563  * Clear the blocking lock, the race is possible between ldlm_handle_convert0()
1564  * and ldlm_work_bl_ast_lock(), so this is done under lock with check for NULL.
1565  */
1566 void ldlm_clear_blocking_lock(struct ldlm_lock *lock)
1567 {
1568         if (lock->l_blocking_lock) {
1569                 LDLM_LOCK_RELEASE(lock->l_blocking_lock);
1570                 lock->l_blocking_lock = NULL;
1571         }
1572 }
1573
1574 /* A lock can be converted to new ibits or mode and should be considered
1575  * as new lock. Clear all states related to a previous blocking AST
1576  * processing so new conflicts will cause new blocking ASTs.
1577  *
1578  * This is used during lock convert below and lock downgrade to COS mode in
1579  * ldlm_lock_mode_downgrade().
1580  */
1581 void ldlm_clear_blocking_data(struct ldlm_lock *lock)
1582 {
1583         ldlm_clear_ast_sent(lock);
1584         lock->l_bl_ast_run = 0;
1585         ldlm_clear_blocking_lock(lock);
1586 }
1587
1588 /**
1589  * Main LDLM entry point for server code to process lock conversion requests.
1590  */
1591 int ldlm_handle_convert0(struct ptlrpc_request *req,
1592                          const struct ldlm_request *dlm_req)
1593 {
1594         struct obd_export *exp = req->rq_export;
1595         struct ldlm_reply *dlm_rep;
1596         struct ldlm_lock *lock;
1597         __u64 bits;
1598         __u64 new_bits;
1599         int rc;
1600
1601         ENTRY;
1602
1603         if (exp && exp->exp_nid_stats && exp->exp_nid_stats->nid_ldlm_stats)
1604                 lprocfs_counter_incr(exp->exp_nid_stats->nid_ldlm_stats,
1605                                      LDLM_CONVERT - LDLM_FIRST_OPC);
1606
1607         rc = req_capsule_server_pack(&req->rq_pill);
1608         if (rc)
1609                 RETURN(rc);
1610
1611         dlm_rep = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
1612         dlm_rep->lock_flags = dlm_req->lock_flags;
1613
1614         lock = ldlm_handle2lock(&dlm_req->lock_handle[0]);
1615         if (!lock) {
1616                 LDLM_DEBUG_NOLOCK("server lock is canceled already");
1617                 req->rq_status = ELDLM_NO_LOCK_DATA;
1618                 RETURN(0);
1619         }
1620
1621         LDLM_DEBUG(lock, "server-side convert handler START");
1622
1623         lock_res_and_lock(lock);
1624         bits = lock->l_policy_data.l_inodebits.bits;
1625         new_bits = dlm_req->lock_desc.l_policy_data.l_inodebits.bits;
1626
1627         if (ldlm_is_cancel(lock)) {
1628                 LDLM_DEBUG(lock, "convert on canceled lock!");
1629                 unlock_res_and_lock(lock);
1630                 GOTO(out_put, rc = ELDLM_NO_LOCK_DATA);
1631         }
1632
1633         if (dlm_req->lock_desc.l_req_mode != lock->l_granted_mode) {
1634                 LDLM_ERROR(lock, "lock mode differs!");
1635                 unlock_res_and_lock(lock);
1636                 GOTO(out_put, rc = -EPROTO);
1637         }
1638
1639         if (bits == new_bits) {
1640                 /*
1641                  * This can be valid situation if CONVERT RPCs are
1642                  * re-ordered. Just finish silently
1643                  */
1644                 LDLM_DEBUG(lock, "lock is converted already!");
1645                 unlock_res_and_lock(lock);
1646         } else {
1647                 if (ldlm_is_waited(lock))
1648                         ldlm_del_waiting_lock(lock);
1649
1650                 ldlm_clear_cbpending(lock);
1651                 lock->l_policy_data.l_inodebits.cancel_bits = 0;
1652                 ldlm_inodebits_drop(lock, bits & ~new_bits);
1653
1654                 ldlm_clear_blocking_data(lock);
1655                 unlock_res_and_lock(lock);
1656
1657                 ldlm_reprocess_all(lock->l_resource, NULL);
1658         }
1659
1660         dlm_rep->lock_handle = lock->l_remote_handle;
1661         ldlm_ibits_policy_local_to_wire(&lock->l_policy_data,
1662                                         &dlm_rep->lock_desc.l_policy_data);
1663         rc = ELDLM_OK;
1664         EXIT;
1665 out_put:
1666         LDLM_DEBUG(lock, "server-side convert handler END, rc = %d", rc);
1667         LDLM_LOCK_PUT(lock);
1668         req->rq_status = rc;
1669         return 0;
1670 }
1671
1672 /**
1673  * Cancel all the locks whose handles are packed into ldlm_request
1674  *
1675  * Called by server code expecting such combined cancel activity
1676  * requests.
1677  */
1678 int ldlm_request_cancel(struct ptlrpc_request *req,
1679                         const struct ldlm_request *dlm_req,
1680                         int first, enum lustre_at_flags flags)
1681 {
1682         struct ldlm_resource *res, *pres = NULL;
1683         struct ldlm_lock *lock;
1684         int i, count, done = 0;
1685         unsigned int size;
1686
1687         ENTRY;
1688
1689         size = req_capsule_get_size(&req->rq_pill, &RMF_DLM_REQ, RCL_CLIENT);
1690         if (size <= offsetof(struct ldlm_request, lock_handle) ||
1691             (size - offsetof(struct ldlm_request, lock_handle)) /
1692              sizeof(struct lustre_handle) < dlm_req->lock_count)
1693                 RETURN(0);
1694
1695         count = dlm_req->lock_count ? dlm_req->lock_count : 1;
1696         if (first >= count)
1697                 RETURN(0);
1698
1699         if (count == 1 && dlm_req->lock_handle[0].cookie == 0)
1700                 RETURN(0);
1701
1702         /*
1703          * There is no lock on the server at the replay time,
1704          * skip lock cancelling to make replay tests to pass.
1705          */
1706         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
1707                 RETURN(0);
1708
1709         LDLM_DEBUG_NOLOCK("server-side cancel handler START: %d locks, starting at %d",
1710                           count, first);
1711
1712         for (i = first; i < count; i++) {
1713                 lock = ldlm_handle2lock(&dlm_req->lock_handle[i]);
1714                 if (!lock) {
1715                         /* below message checked in replay-single.sh test_36 */
1716                         LDLM_DEBUG_NOLOCK("server-side cancel handler stale lock (cookie %llu)",
1717                                           dlm_req->lock_handle[i].cookie);
1718                         continue;
1719                 }
1720
1721                 res = lock->l_resource;
1722                 done++;
1723
1724                 /*
1725                  * This code is an optimization to only attempt lock
1726                  * granting on the resource (that could be CPU-expensive)
1727                  * after we are done cancelling lock in that resource.
1728                  */
1729                 if (res != pres) {
1730                         if (pres != NULL) {
1731                                 ldlm_reprocess_all(pres, NULL);
1732                                 LDLM_RESOURCE_DELREF(pres);
1733                                 ldlm_resource_putref(pres);
1734                         }
1735                         if (res != NULL) {
1736                                 ldlm_resource_getref(res);
1737                                 LDLM_RESOURCE_ADDREF(res);
1738
1739                                 if (!ldlm_is_discard_data(lock))
1740                                         ldlm_lvbo_update(res, lock,
1741                                                          NULL, 1);
1742                         }
1743                         pres = res;
1744                 }
1745
1746                 if ((flags & LATF_STATS) && ldlm_is_ast_sent(lock) &&
1747                     lock->l_blast_sent != 0) {
1748                         timeout_t delay = 0;
1749
1750                         if (ktime_get_real_seconds() > lock->l_blast_sent)
1751                                 delay = ktime_get_real_seconds() -
1752                                         lock->l_blast_sent;
1753                         LDLM_DEBUG(lock,
1754                                    "server cancels blocked lock after %ds",
1755                                    delay);
1756                         at_measured(&lock->l_export->exp_bl_lock_at, delay);
1757                 }
1758                 ldlm_lock_cancel(lock);
1759                 LDLM_LOCK_PUT(lock);
1760         }
1761         if (pres != NULL) {
1762                 ldlm_reprocess_all(pres, NULL);
1763                 LDLM_RESOURCE_DELREF(pres);
1764                 ldlm_resource_putref(pres);
1765         }
1766         LDLM_DEBUG_NOLOCK("server-side cancel handler END");
1767         RETURN(done);
1768 }
1769 EXPORT_SYMBOL(ldlm_request_cancel);
1770
1771 /**
1772  * Main LDLM entry point for server code to cancel locks.
1773  *
1774  * Typically gets called from service handler on LDLM_CANCEL opc.
1775  */
1776 int ldlm_handle_cancel(struct ptlrpc_request *req)
1777 {
1778         struct ldlm_request *dlm_req;
1779         int rc;
1780
1781         ENTRY;
1782
1783         dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1784         if (dlm_req == NULL) {
1785                 CDEBUG(D_INFO, "bad request buffer for cancel\n");
1786                 RETURN(-EFAULT);
1787         }
1788
1789         if (req_capsule_get_size(&req->rq_pill, &RMF_DLM_REQ, RCL_CLIENT) <
1790             offsetof(struct ldlm_request, lock_handle[1]))
1791                 RETURN(-EPROTO);
1792
1793         if (req->rq_export && req->rq_export->exp_nid_stats &&
1794             req->rq_export->exp_nid_stats->nid_ldlm_stats)
1795                 lprocfs_counter_incr(req->rq_export->exp_nid_stats->nid_ldlm_stats,
1796                                      LDLM_CANCEL - LDLM_FIRST_OPC);
1797
1798         rc = req_capsule_server_pack(&req->rq_pill);
1799         if (rc)
1800                 RETURN(rc);
1801
1802         if (!ldlm_request_cancel(req, dlm_req, 0, LATF_STATS))
1803                 req->rq_status = LUSTRE_ESTALE;
1804
1805         RETURN(ptlrpc_reply(req));
1806 }
1807 #endif /* HAVE_SERVER_SUPPORT */
1808
1809 /**
1810  * Server may pass additional information about blocking lock.
1811  * For IBITS locks it is conflicting bits which can be used for
1812  * lock convert instead of cancel.
1813  */
1814 void ldlm_bl_desc2lock(const struct ldlm_lock_desc *ld, struct ldlm_lock *lock)
1815 {
1816         struct ldlm_namespace *ns = ldlm_lock_to_ns(lock);
1817
1818         check_res_locked(lock->l_resource);
1819         if (ns_is_client(ns) && ld &&
1820             (lock->l_resource->lr_type == LDLM_IBITS)) {
1821                 /*
1822                  * Lock description contains policy of blocking lock,
1823                  * and its cancel_bits is used to pass conflicting bits.
1824                  * NOTE: ld can be NULL or can be not NULL but zeroed if
1825                  * passed from ldlm_bl_thread_blwi(), check below used bits
1826                  * in ld to make sure it is valid description.
1827                  *
1828                  * If server may replace lock resource keeping the same cookie,
1829                  * never use cancel bits from different resource, full cancel
1830                  * is to be used.
1831                  */
1832                 if (ld->l_policy_data.l_inodebits.cancel_bits &&
1833                     ldlm_res_eq(&ld->l_resource.lr_name,
1834                                 &lock->l_resource->lr_name) &&
1835                     !(ldlm_is_cbpending(lock) &&
1836                       lock->l_policy_data.l_inodebits.cancel_bits == 0)) {
1837                         /* always combine conflicting ibits */
1838                         lock->l_policy_data.l_inodebits.cancel_bits |=
1839                                 ld->l_policy_data.l_inodebits.cancel_bits;
1840                 } else {
1841                         /* If cancel_bits are not obtained or
1842                          * if the lock is already CBPENDING and
1843                          * has no cancel_bits set
1844                          * - the full lock is to be cancelled
1845                          */
1846                         lock->l_policy_data.l_inodebits.cancel_bits = 0;
1847                 }
1848         }
1849 }
1850
1851 /**
1852  * Callback handler for receiving incoming blocking ASTs.
1853  *
1854  * This can only happen on client side.
1855  */
1856 void ldlm_handle_bl_callback(struct ldlm_namespace *ns,
1857                              struct ldlm_lock_desc *ld, struct ldlm_lock *lock)
1858 {
1859         int do_ast;
1860
1861         ENTRY;
1862
1863         LDLM_DEBUG(lock, "client blocking AST callback handler");
1864
1865         lock_res_and_lock(lock);
1866
1867         /* get extra information from desc if any */
1868         ldlm_bl_desc2lock(ld, lock);
1869         ldlm_set_cbpending(lock);
1870
1871         do_ast = (!lock->l_readers && !lock->l_writers);
1872         unlock_res_and_lock(lock);
1873
1874         if (do_ast) {
1875                 CDEBUG(D_DLMTRACE,
1876                        "Lock %p already unused, calling callback (%p)\n",
1877                        lock, lock->l_blocking_ast);
1878                 if (lock->l_blocking_ast != NULL)
1879                         lock->l_blocking_ast(lock, ld, lock->l_ast_data,
1880                                              LDLM_CB_BLOCKING);
1881         } else {
1882                 CDEBUG(D_DLMTRACE,
1883                        "Lock %p is referenced, will be cancelled later\n",
1884                        lock);
1885         }
1886
1887         LDLM_DEBUG(lock, "client blocking callback handler END");
1888         LDLM_LOCK_RELEASE(lock);
1889         EXIT;
1890 }
1891
1892 static int ldlm_callback_reply(struct ptlrpc_request *req, int rc)
1893 {
1894         if (req->rq_no_reply)
1895                 return 0;
1896
1897         req->rq_status = rc;
1898         if (!req->rq_packed_final) {
1899                 rc = lustre_pack_reply(req, 1, NULL, NULL);
1900                 if (rc)
1901                         return rc;
1902         }
1903         return ptlrpc_reply(req);
1904 }
1905
1906 /**
1907  * Callback handler for receiving incoming completion ASTs.
1908  *
1909  * This only can happen on client side.
1910  */
1911 static int ldlm_handle_cp_callback(struct ptlrpc_request *req,
1912                                     struct ldlm_namespace *ns,
1913                                     struct ldlm_request *dlm_req,
1914                                     struct ldlm_lock *lock)
1915 {
1916         LIST_HEAD(ast_list);
1917         int lvb_len;
1918         int rc = 0;
1919
1920         ENTRY;
1921
1922         LDLM_DEBUG(lock, "client completion callback handler START");
1923
1924         if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_BL_CB_RACE)) {
1925                 long to = cfs_time_seconds(1);
1926
1927                 ldlm_callback_reply(req, 0);
1928
1929                 while (to > 0) {
1930                         schedule_timeout_interruptible(to);
1931                         if (ldlm_is_granted(lock) ||
1932                             ldlm_is_destroyed(lock))
1933                                 break;
1934                 }
1935         }
1936
1937         lvb_len = req_capsule_get_size(&req->rq_pill, &RMF_DLM_LVB, RCL_CLIENT);
1938         if (lvb_len < 0) {
1939                 LDLM_ERROR(lock, "Fail to get lvb_len, rc = %d", lvb_len);
1940                 GOTO(out, rc = lvb_len);
1941         } else if (lvb_len > 0) {
1942                 if (lock->l_lvb_len > 0) {
1943                         /* for extent lock, lvb contains ost_lvb{}. */
1944                         LASSERT(lock->l_lvb_data != NULL);
1945
1946                         if (unlikely(lock->l_lvb_len < lvb_len)) {
1947                                 LDLM_ERROR(lock,
1948                                            "Replied LVB is larger than expectation, expected = %d, replied = %d",
1949                                            lock->l_lvb_len, lvb_len);
1950                                 GOTO(out, rc = -EINVAL);
1951                         }
1952                 }
1953         }
1954
1955         lock_res_and_lock(lock);
1956
1957         if (!ldlm_res_eq(&dlm_req->lock_desc.l_resource.lr_name,
1958                          &lock->l_resource->lr_name)) {
1959                 ldlm_resource_unlink_lock(lock);
1960                 unlock_res_and_lock(lock);
1961                 rc = ldlm_lock_change_resource(ns, lock,
1962                                 &dlm_req->lock_desc.l_resource.lr_name);
1963                 if (rc < 0) {
1964                         LDLM_ERROR(lock, "Failed to allocate resource");
1965                         GOTO(out, rc);
1966                 }
1967                 LDLM_DEBUG(lock, "completion AST, new resource");
1968                 lock_res_and_lock(lock);
1969         }
1970
1971         if (ldlm_is_failed(lock)) {
1972                 unlock_res_and_lock(lock);
1973                 LDLM_LOCK_RELEASE(lock);
1974                 RETURN(-EINVAL);
1975         }
1976
1977         if (ldlm_is_destroyed(lock) ||
1978             ldlm_is_granted(lock)) {
1979                 /* b=11300: the lock has already been granted */
1980                 unlock_res_and_lock(lock);
1981                 LDLM_DEBUG(lock, "Double grant race happened");
1982                 GOTO(out, rc = 0);
1983         }
1984
1985         /*
1986          * If we receive the completion AST before the actual enqueue returned,
1987          * then we might need to switch lock modes, resources, or extents.
1988          */
1989         if (dlm_req->lock_desc.l_granted_mode != lock->l_req_mode) {
1990                 lock->l_req_mode = dlm_req->lock_desc.l_granted_mode;
1991                 LDLM_DEBUG(lock, "completion AST, new lock mode");
1992         }
1993
1994         if (lock->l_resource->lr_type != LDLM_PLAIN) {
1995                 ldlm_convert_policy_to_local(req->rq_export,
1996                                           dlm_req->lock_desc.l_resource.lr_type,
1997                                           &dlm_req->lock_desc.l_policy_data,
1998                                           &lock->l_policy_data);
1999                 LDLM_DEBUG(lock, "completion AST, new policy data");
2000         }
2001
2002         ldlm_resource_unlink_lock(lock);
2003
2004         if (dlm_req->lock_flags & LDLM_FL_AST_SENT) {
2005                 /*
2006                  * BL_AST locks are not needed in LRU.
2007                  * Let ldlm_cancel_lru() be fast.
2008                  */
2009                 ldlm_lock_remove_from_lru(lock);
2010                 ldlm_bl_desc2lock(&dlm_req->lock_desc, lock);
2011                 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_BL_AST;
2012                 LDLM_DEBUG(lock, "completion AST includes blocking AST");
2013         }
2014
2015         if (lock->l_lvb_len > 0) {
2016                 rc = ldlm_fill_lvb(lock, &req->rq_pill, RCL_CLIENT,
2017                                    lock->l_lvb_data, lvb_len);
2018                 if (rc < 0) {
2019                         unlock_res_and_lock(lock);
2020                         GOTO(out, rc);
2021                 }
2022         }
2023
2024         ldlm_grant_lock(lock, &ast_list);
2025         unlock_res_and_lock(lock);
2026
2027         LDLM_DEBUG(lock, "callback handler finished, about to run_ast_work");
2028
2029         /*
2030          * Let Enqueue to call osc_lock_upcall() and initialize
2031          * l_ast_data
2032          */
2033         OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_CP_ENQ_RACE, 2);
2034
2035         ldlm_run_ast_work(ns, &ast_list, LDLM_WORK_CP_AST);
2036
2037         LDLM_DEBUG_NOLOCK("client completion callback handler END (lock %p)",
2038                           lock);
2039         GOTO(out, rc);
2040
2041 out:
2042         if (rc < 0) {
2043                 lock_res_and_lock(lock);
2044                 ldlm_set_failed(lock);
2045                 unlock_res_and_lock(lock);
2046                 wake_up(&lock->l_waitq);
2047         }
2048         LDLM_LOCK_RELEASE(lock);
2049
2050         return 0;
2051 }
2052
2053 /**
2054  * Callback handler for receiving incoming glimpse ASTs.
2055  *
2056  * This only can happen on client side.  After handling the glimpse AST
2057  * we also consider dropping the lock here if it is unused locally for a
2058  * long time.
2059  */
2060 static void ldlm_handle_gl_callback(struct ptlrpc_request *req,
2061                                     struct ldlm_namespace *ns,
2062                                     struct ldlm_request *dlm_req,
2063                                     struct ldlm_lock *lock)
2064 {
2065         struct ldlm_lock_desc *ld = &dlm_req->lock_desc;
2066         int rc = -ENOSYS;
2067
2068         ENTRY;
2069
2070         LDLM_DEBUG(lock, "client glimpse AST callback handler");
2071
2072         if (lock->l_glimpse_ast != NULL)
2073                 rc = lock->l_glimpse_ast(lock, req);
2074
2075         if (req->rq_repmsg != NULL) {
2076                 ptlrpc_reply(req);
2077         } else {
2078                 req->rq_status = rc;
2079                 ptlrpc_error(req);
2080         }
2081
2082         lock_res_and_lock(lock);
2083         if (lock->l_granted_mode == LCK_PW &&
2084             !lock->l_readers && !lock->l_writers &&
2085             ktime_after(ktime_get(),
2086                         ktime_add(lock->l_last_used, ns->ns_dirty_age_limit))) {
2087                 unlock_res_and_lock(lock);
2088
2089                 /* For MDS glimpse it is always DOM lock, set corresponding
2090                  * cancel_bits to perform lock convert if needed
2091                  */
2092                 if (lock->l_resource->lr_type == LDLM_IBITS)
2093                         ld->l_policy_data.l_inodebits.cancel_bits =
2094                                                         MDS_INODELOCK_DOM;
2095                 if (ldlm_bl_to_thread_lock(ns, ld, lock))
2096                         ldlm_handle_bl_callback(ns, ld, lock);
2097
2098                 EXIT;
2099                 return;
2100         }
2101         unlock_res_and_lock(lock);
2102         LDLM_LOCK_RELEASE(lock);
2103         EXIT;
2104 }
2105
2106 static int __ldlm_bl_to_thread(struct ldlm_bl_work_item *blwi,
2107                                enum ldlm_cancel_flags cancel_flags)
2108 {
2109         struct ldlm_bl_pool *blp = ldlm_state->ldlm_bl_pool;
2110
2111         ENTRY;
2112
2113         spin_lock(&blp->blp_lock);
2114         if (blwi->blwi_lock &&
2115             ldlm_is_discard_data(blwi->blwi_lock)) {
2116                 /* add LDLM_FL_DISCARD_DATA requests to the priority list */
2117                 list_add_tail(&blwi->blwi_entry, &blp->blp_prio_list);
2118         } else {
2119                 /* other blocking callbacks are added to the regular list */
2120                 list_add_tail(&blwi->blwi_entry, &blp->blp_list);
2121         }
2122         spin_unlock(&blp->blp_lock);
2123
2124         wake_up(&blp->blp_waitq);
2125
2126         /*
2127          * can not check blwi->blwi_flags as blwi could be already freed in
2128          * LCF_ASYNC mode
2129          */
2130         if (!(cancel_flags & LCF_ASYNC))
2131                 wait_for_completion(&blwi->blwi_comp);
2132
2133         RETURN(0);
2134 }
2135
2136 static inline void init_blwi(struct ldlm_bl_work_item *blwi,
2137                              struct ldlm_namespace *ns,
2138                              struct ldlm_lock_desc *ld,
2139                              struct list_head *cancels, int count,
2140                              struct ldlm_lock *lock,
2141                              enum ldlm_cancel_flags cancel_flags)
2142 {
2143         init_completion(&blwi->blwi_comp);
2144         INIT_LIST_HEAD(&blwi->blwi_head);
2145
2146         if (current->flags & PF_MEMALLOC)
2147                 blwi->blwi_mem_pressure = 1;
2148
2149         blwi->blwi_ns = ns;
2150         blwi->blwi_flags = cancel_flags;
2151         if (ld != NULL)
2152                 blwi->blwi_ld = *ld;
2153         if (count) {
2154                 list_splice_init(cancels, &blwi->blwi_head);
2155                 blwi->blwi_count = count;
2156         } else {
2157                 blwi->blwi_lock = lock;
2158         }
2159 }
2160
2161 /**
2162  * Queues a list of locks \a cancels containing \a count locks
2163  * for later processing by a blocking thread.  If \a count is zero,
2164  * then the lock referenced as \a lock is queued instead.
2165  *
2166  * The blocking thread would then call ->l_blocking_ast callback in the lock.
2167  * If list addition fails an error is returned and caller is supposed to
2168  * call ->l_blocking_ast itself.
2169  */
2170 static int ldlm_bl_to_thread(struct ldlm_namespace *ns,
2171                              struct ldlm_lock_desc *ld,
2172                              struct ldlm_lock *lock,
2173                              struct list_head *cancels, int count,
2174                              enum ldlm_cancel_flags cancel_flags)
2175 {
2176         ENTRY;
2177
2178         if (cancels && count == 0)
2179                 RETURN(0);
2180
2181         if (cancel_flags & LCF_ASYNC) {
2182                 struct ldlm_bl_work_item *blwi;
2183
2184                 OBD_ALLOC(blwi, sizeof(*blwi));
2185                 if (blwi == NULL)
2186                         RETURN(-ENOMEM);
2187                 init_blwi(blwi, ns, ld, cancels, count, lock, cancel_flags);
2188
2189                 RETURN(__ldlm_bl_to_thread(blwi, cancel_flags));
2190         } else {
2191                 /*
2192                  * if it is synchronous call do minimum mem alloc, as it could
2193                  * be triggered from kernel shrinker
2194                  */
2195                 struct ldlm_bl_work_item blwi;
2196
2197                 memset(&blwi, 0, sizeof(blwi));
2198                 init_blwi(&blwi, ns, ld, cancels, count, lock, cancel_flags);
2199                 RETURN(__ldlm_bl_to_thread(&blwi, cancel_flags));
2200         }
2201 }
2202
2203
2204 int ldlm_bl_to_thread_lock(struct ldlm_namespace *ns, struct ldlm_lock_desc *ld,
2205                            struct ldlm_lock *lock)
2206 {
2207         return ldlm_bl_to_thread(ns, ld, lock, NULL, 0, LCF_ASYNC);
2208 }
2209
2210 int ldlm_bl_to_thread_list(struct ldlm_namespace *ns, struct ldlm_lock_desc *ld,
2211                            struct list_head *cancels, int count,
2212                            enum ldlm_cancel_flags cancel_flags)
2213 {
2214         return ldlm_bl_to_thread(ns, ld, NULL, cancels, count, cancel_flags);
2215 }
2216
2217 int ldlm_bl_thread_wakeup(void)
2218 {
2219         wake_up(&ldlm_state->ldlm_bl_pool->blp_waitq);
2220         return 0;
2221 }
2222
2223 /* Setinfo coming from Server (eg MDT) to Client (eg MDC)! */
2224 static int ldlm_handle_setinfo(struct ptlrpc_request *req)
2225 {
2226         struct obd_device *obd = req->rq_export->exp_obd;
2227         char *key;
2228         void *val;
2229         int keylen, vallen;
2230         int rc = -ENOSYS;
2231
2232         ENTRY;
2233
2234         DEBUG_REQ(D_HSM, req, "%s: handle setinfo", obd->obd_name);
2235
2236         req_capsule_set(&req->rq_pill, &RQF_OBD_SET_INFO);
2237
2238         key = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_KEY);
2239         if (key == NULL) {
2240                 DEBUG_REQ(D_IOCTL, req, "no set_info key");
2241                 RETURN(-EFAULT);
2242         }
2243         keylen = req_capsule_get_size(&req->rq_pill, &RMF_SETINFO_KEY,
2244                                       RCL_CLIENT);
2245         val = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_VAL);
2246         if (val == NULL) {
2247                 DEBUG_REQ(D_IOCTL, req, "no set_info val");
2248                 RETURN(-EFAULT);
2249         }
2250         vallen = req_capsule_get_size(&req->rq_pill, &RMF_SETINFO_VAL,
2251                                       RCL_CLIENT);
2252
2253         /* We are responsible for swabbing contents of val */
2254
2255         if (KEY_IS(KEY_HSM_COPYTOOL_SEND))
2256                 /* Pass it on to mdc (the "export" in this case) */
2257                 rc = obd_set_info_async(req->rq_svc_thread->t_env,
2258                                         req->rq_export,
2259                                         sizeof(KEY_HSM_COPYTOOL_SEND),
2260                                         KEY_HSM_COPYTOOL_SEND,
2261                                         vallen, val, NULL);
2262         else
2263                 DEBUG_REQ(D_WARNING, req, "ignoring unknown key '%s'", key);
2264
2265         return rc;
2266 }
2267
2268 static inline void ldlm_callback_errmsg(struct ptlrpc_request *req,
2269                                         const char *msg, int rc,
2270                                         const struct lustre_handle *handle)
2271 {
2272         DEBUG_REQ((req->rq_no_reply || rc) ? D_WARNING : D_DLMTRACE, req,
2273                   "%s, NID=%s lock=%#llx: rc = %d",
2274                   msg, libcfs_id2str(req->rq_peer),
2275                   handle ? handle->cookie : 0, rc);
2276         if (req->rq_no_reply)
2277                 CWARN("No reply was sent, maybe cause b=21636.\n");
2278         else if (rc)
2279                 CWARN("Send reply failed, maybe cause b=21636.\n");
2280 }
2281
2282 /* TODO: handle requests in a similar way as MDT: see mdt_handle_common() */
2283 static int ldlm_callback_handler(struct ptlrpc_request *req)
2284 {
2285         struct ldlm_namespace *ns;
2286         struct ldlm_request *dlm_req;
2287         struct ldlm_lock *lock;
2288         int rc;
2289
2290         ENTRY;
2291
2292         /*
2293          * Requests arrive in sender's byte order.  The ptlrpc service
2294          * handler has already checked and, if necessary, byte-swapped the
2295          * incoming request message body, but I am responsible for the
2296          * message buffers.
2297          */
2298
2299         /* do nothing for sec context finalize */
2300         if (lustre_msg_get_opc(req->rq_reqmsg) == SEC_CTX_FINI)
2301                 RETURN(0);
2302
2303         req_capsule_init(&req->rq_pill, req, RCL_SERVER);
2304
2305         if (req->rq_export == NULL) {
2306                 rc = ldlm_callback_reply(req, -ENOTCONN);
2307                 ldlm_callback_errmsg(req, "Operate on unconnected server",
2308                                      rc, NULL);
2309                 RETURN(0);
2310         }
2311
2312         LASSERT(req->rq_export != NULL);
2313         LASSERT(req->rq_export->exp_obd != NULL);
2314
2315         switch (lustre_msg_get_opc(req->rq_reqmsg)) {
2316         case LDLM_BL_CALLBACK:
2317                 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_BL_CALLBACK_NET)) {
2318                         if (cfs_fail_err)
2319                                 ldlm_callback_reply(req, -(int)cfs_fail_err);
2320                         RETURN(0);
2321                 }
2322                 break;
2323         case LDLM_CP_CALLBACK:
2324                 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CP_CALLBACK_NET))
2325                         RETURN(0);
2326                 break;
2327         case LDLM_GL_CALLBACK:
2328                 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_GL_CALLBACK_NET))
2329                         RETURN(0);
2330                 break;
2331         case LDLM_SET_INFO:
2332                 rc = ldlm_handle_setinfo(req);
2333                 ldlm_callback_reply(req, rc);
2334                 RETURN(0);
2335         default:
2336                 CERROR("unknown opcode %u\n",
2337                        lustre_msg_get_opc(req->rq_reqmsg));
2338                 ldlm_callback_reply(req, -EPROTO);
2339                 RETURN(0);
2340         }
2341
2342         ns = req->rq_export->exp_obd->obd_namespace;
2343         LASSERT(ns != NULL);
2344
2345         req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK);
2346
2347         dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2348         if (dlm_req == NULL) {
2349                 rc = ldlm_callback_reply(req, -EPROTO);
2350                 ldlm_callback_errmsg(req, "Operate without parameter", rc,
2351                                      NULL);
2352                 RETURN(0);
2353         }
2354
2355         /*
2356          * Force a known safe race, send a cancel to the server for a lock
2357          * which the server has already started a blocking callback on.
2358          */
2359         if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_BL_CB_RACE) &&
2360             lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK) {
2361                 rc = ldlm_cli_cancel(&dlm_req->lock_handle[0], 0);
2362                 if (rc < 0)
2363                         CERROR("ldlm_cli_cancel: %d\n", rc);
2364         }
2365
2366         lock = ldlm_handle2lock_long(&dlm_req->lock_handle[0], 0);
2367         if (!lock) {
2368                 CDEBUG(D_DLMTRACE,
2369                        "callback on lock %#llx - lock disappeared\n",
2370                        dlm_req->lock_handle[0].cookie);
2371                 rc = ldlm_callback_reply(req, -EINVAL);
2372                 ldlm_callback_errmsg(req, "Operate with invalid parameter", rc,
2373                                      &dlm_req->lock_handle[0]);
2374                 RETURN(0);
2375         }
2376
2377         if (ldlm_is_fail_loc(lock) &&
2378             lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK)
2379                 OBD_RACE(OBD_FAIL_LDLM_CP_BL_RACE);
2380
2381         /* Copy hints/flags (e.g. LDLM_FL_DISCARD_DATA) from AST. */
2382         lock_res_and_lock(lock);
2383         lock->l_flags |= ldlm_flags_from_wire(dlm_req->lock_flags &
2384                                               LDLM_FL_AST_MASK);
2385         if (lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK) {
2386                 /*
2387                  * If somebody cancels lock and cache is already dropped,
2388                  * or lock is failed before cp_ast received on client,
2389                  * we can tell the server we have no lock. Otherwise, we
2390                  * should send cancel after dropping the cache.
2391                  */
2392                 if ((ldlm_is_canceling(lock) && ldlm_is_bl_done(lock)) ||
2393                      ldlm_is_failed(lock)) {
2394                         LDLM_DEBUG(lock,
2395                                    "callback on lock %llx - lock disappeared",
2396                                    dlm_req->lock_handle[0].cookie);
2397                         unlock_res_and_lock(lock);
2398                         LDLM_LOCK_RELEASE(lock);
2399                         rc = ldlm_callback_reply(req, -EINVAL);
2400                         ldlm_callback_errmsg(req, "Operate on stale lock", rc,
2401                                              &dlm_req->lock_handle[0]);
2402                         RETURN(0);
2403                 }
2404                 /*
2405                  * BL_AST locks are not needed in LRU.
2406                  * Let ldlm_cancel_lru() be fast.
2407                  */
2408                 ldlm_lock_remove_from_lru(lock);
2409                 ldlm_set_bl_ast(lock);
2410         }
2411         unlock_res_and_lock(lock);
2412
2413         /*
2414          * We want the ost thread to get this reply so that it can respond
2415          * to ost requests (write cache writeback) that might be triggered
2416          * in the callback.
2417          *
2418          * But we'd also like to be able to indicate in the reply that we're
2419          * cancelling right now, because it's unused, or have an intent result
2420          * in the reply, so we might have to push the responsibility for sending
2421          * the reply down into the AST handlers, alas.
2422          */
2423
2424         switch (lustre_msg_get_opc(req->rq_reqmsg)) {
2425         case LDLM_BL_CALLBACK:
2426                 CDEBUG(D_INODE, "blocking ast\n");
2427                 req_capsule_extend(&req->rq_pill, &RQF_LDLM_BL_CALLBACK);
2428                 if (!ldlm_is_cancel_on_block(lock)) {
2429                         rc = ldlm_callback_reply(req, 0);
2430                         if (req->rq_no_reply || rc)
2431                                 ldlm_callback_errmsg(req, "Normal process", rc,
2432                                                      &dlm_req->lock_handle[0]);
2433                 }
2434                 if (ldlm_bl_to_thread_lock(ns, &dlm_req->lock_desc, lock))
2435                         ldlm_handle_bl_callback(ns, &dlm_req->lock_desc, lock);
2436                 break;
2437         case LDLM_CP_CALLBACK:
2438                 CDEBUG(D_INODE, "completion ast\n");
2439                 req_capsule_extend(&req->rq_pill, &RQF_LDLM_CP_CALLBACK);
2440                 rc = ldlm_handle_cp_callback(req, ns, dlm_req, lock);
2441                 if (!OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_BL_CB_RACE))
2442                         ldlm_callback_reply(req, rc);
2443                 break;
2444         case LDLM_GL_CALLBACK:
2445                 CDEBUG(D_INODE, "glimpse ast\n");
2446                 req_capsule_extend(&req->rq_pill, &RQF_LDLM_GL_CALLBACK);
2447                 ldlm_handle_gl_callback(req, ns, dlm_req, lock);
2448                 break;
2449         default:
2450                 LBUG(); /* checked above */
2451         }
2452
2453         RETURN(0);
2454 }
2455
2456 #ifdef HAVE_SERVER_SUPPORT
2457 /**
2458  * Main handler for canceld thread.
2459  *
2460  * Separated into its own thread to avoid deadlocks.
2461  */
2462 static int ldlm_cancel_handler(struct ptlrpc_request *req)
2463 {
2464         int rc;
2465
2466         ENTRY;
2467
2468         /*
2469          * Requests arrive in sender's byte order.  The ptlrpc service
2470          * handler has already checked and, if necessary, byte-swapped the
2471          * incoming request message body, but I am responsible for the
2472          * message buffers.
2473          */
2474
2475         req_capsule_init(&req->rq_pill, req, RCL_SERVER);
2476
2477         if (req->rq_export == NULL) {
2478                 struct ldlm_request *dlm_req;
2479
2480                 CERROR("%s from %s arrived at %llu with bad export cookie %llu\n",
2481                        ll_opcode2str(lustre_msg_get_opc(req->rq_reqmsg)),
2482                        libcfs_nid2str(req->rq_peer.nid),
2483                        (unsigned long long)req->rq_arrival_time.tv_sec,
2484                        lustre_msg_get_handle(req->rq_reqmsg)->cookie);
2485
2486                 if (lustre_msg_get_opc(req->rq_reqmsg) == LDLM_CANCEL) {
2487                         req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK);
2488                         dlm_req = req_capsule_client_get(&req->rq_pill,
2489                                                          &RMF_DLM_REQ);
2490                         if (dlm_req != NULL)
2491                                 ldlm_lock_dump_handle(D_ERROR,
2492                                                       &dlm_req->lock_handle[0]);
2493                 }
2494                 ldlm_callback_reply(req, -ENOTCONN);
2495                 RETURN(0);
2496         }
2497
2498         switch (lustre_msg_get_opc(req->rq_reqmsg)) {
2499         /* XXX FIXME move this back to mds/handler.c, b=249 */
2500         case LDLM_CANCEL:
2501                 req_capsule_set(&req->rq_pill, &RQF_LDLM_CANCEL);
2502                 CDEBUG(D_INODE, "cancel\n");
2503                 if (CFS_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_NET) ||
2504                     CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_CANCEL_RESEND) ||
2505                     CFS_FAIL_CHECK(OBD_FAIL_LDLM_BL_EVICT))
2506                         RETURN(0);
2507                 rc = ldlm_handle_cancel(req);
2508                 break;
2509         case LDLM_CONVERT:
2510         {
2511                 struct ldlm_request *dlm_req;
2512
2513                 req_capsule_set(&req->rq_pill, &RQF_LDLM_CONVERT);
2514                 CDEBUG(D_INODE, "convert\n");
2515
2516                 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2517                 if (dlm_req == NULL) {
2518                         CDEBUG(D_INFO, "bad request buffer for cancel\n");
2519                         rc = ldlm_callback_reply(req, -EPROTO);
2520                 } else {
2521                         req->rq_status = ldlm_handle_convert0(req, dlm_req);
2522                         rc = ptlrpc_reply(req);
2523                 }
2524                 break;
2525         }
2526         default:
2527                 CERROR("invalid opcode %d\n",
2528                        lustre_msg_get_opc(req->rq_reqmsg));
2529                 req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK);
2530                 rc = ldlm_callback_reply(req, -EINVAL);
2531         }
2532
2533         RETURN(rc);
2534 }
2535
2536 static int ldlm_cancel_hpreq_lock_match(struct ptlrpc_request *req,
2537                                         struct ldlm_lock *lock)
2538 {
2539         struct ldlm_request *dlm_req;
2540         struct lustre_handle lockh;
2541         int rc = 0;
2542         int i;
2543
2544         ENTRY;
2545
2546         dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2547         if (dlm_req == NULL)
2548                 RETURN(0);
2549
2550         ldlm_lock2handle(lock, &lockh);
2551         for (i = 0; i < dlm_req->lock_count; i++) {
2552                 if (lustre_handle_equal(&dlm_req->lock_handle[i],
2553                                         &lockh)) {
2554                         DEBUG_REQ(D_RPCTRACE, req,
2555                                   "Prio raised by lock %#llx", lockh.cookie);
2556                         rc = 1;
2557                         break;
2558                 }
2559         }
2560
2561         RETURN(rc);
2562 }
2563
2564 static int ldlm_cancel_hpreq_check(struct ptlrpc_request *req)
2565 {
2566         struct ldlm_request *dlm_req;
2567         int rc = 0;
2568         int i;
2569         unsigned int size;
2570
2571         ENTRY;
2572
2573         /* no prolong in recovery */
2574         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
2575                 RETURN(0);
2576
2577         dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2578         if (dlm_req == NULL)
2579                 RETURN(-EFAULT);
2580
2581         size = req_capsule_get_size(&req->rq_pill, &RMF_DLM_REQ, RCL_CLIENT);
2582         if (size <= offsetof(struct ldlm_request, lock_handle) ||
2583             (size - offsetof(struct ldlm_request, lock_handle)) /
2584              sizeof(struct lustre_handle) < dlm_req->lock_count)
2585                 RETURN(-EPROTO);
2586
2587         for (i = 0; i < dlm_req->lock_count; i++) {
2588                 struct ldlm_lock *lock;
2589
2590                 lock = ldlm_handle2lock(&dlm_req->lock_handle[i]);
2591                 if (lock == NULL)
2592                         continue;
2593
2594                 rc = ldlm_is_ast_sent(lock) ? 1 : 0;
2595                 if (rc)
2596                         LDLM_DEBUG(lock, "hpreq cancel/convert lock");
2597                 LDLM_LOCK_PUT(lock);
2598
2599                 if (rc)
2600                         break;
2601         }
2602
2603         RETURN(rc);
2604 }
2605
2606 static struct ptlrpc_hpreq_ops ldlm_cancel_hpreq_ops = {
2607         .hpreq_lock_match = ldlm_cancel_hpreq_lock_match,
2608         .hpreq_check      = ldlm_cancel_hpreq_check,
2609         .hpreq_fini       = NULL,
2610 };
2611
2612 static int ldlm_hpreq_handler(struct ptlrpc_request *req)
2613 {
2614         ENTRY;
2615
2616         req_capsule_init(&req->rq_pill, req, RCL_SERVER);
2617
2618         if (req->rq_export == NULL)
2619                 RETURN(0);
2620
2621         if (LDLM_CANCEL == lustre_msg_get_opc(req->rq_reqmsg)) {
2622                 req_capsule_set(&req->rq_pill, &RQF_LDLM_CANCEL);
2623                 req->rq_ops = &ldlm_cancel_hpreq_ops;
2624         } else if (LDLM_CONVERT == lustre_msg_get_opc(req->rq_reqmsg)) {
2625                 req_capsule_set(&req->rq_pill, &RQF_LDLM_CONVERT);
2626                 req->rq_ops = &ldlm_cancel_hpreq_ops;
2627         }
2628         RETURN(0);
2629 }
2630
2631 static int ldlm_revoke_lock_cb(struct cfs_hash *hs, struct cfs_hash_bd *bd,
2632                                struct hlist_node *hnode, void *data)
2633
2634 {
2635         struct list_head *rpc_list = data;
2636         struct ldlm_lock *lock = cfs_hash_object(hs, hnode);
2637
2638         lock_res_and_lock(lock);
2639
2640         if (!ldlm_is_granted(lock)) {
2641                 unlock_res_and_lock(lock);
2642                 return 0;
2643         }
2644
2645         LASSERT(lock->l_resource);
2646         if (lock->l_resource->lr_type != LDLM_IBITS &&
2647             lock->l_resource->lr_type != LDLM_PLAIN) {
2648                 unlock_res_and_lock(lock);
2649                 return 0;
2650         }
2651
2652         if (ldlm_is_ast_sent(lock)) {
2653                 unlock_res_and_lock(lock);
2654                 return 0;
2655         }
2656
2657         LASSERT(lock->l_blocking_ast);
2658         LASSERT(!lock->l_blocking_lock);
2659
2660         ldlm_set_ast_sent(lock);
2661         if (lock->l_export && lock->l_export->exp_lock_hash) {
2662                 /*
2663                  * NB: it's safe to call cfs_hash_del() even lock isn't
2664                  * in exp_lock_hash.
2665                  */
2666                 /*
2667                  * In the function below, .hs_keycmp resolves to
2668                  * ldlm_export_lock_keycmp()
2669                  */
2670                 /* coverity[overrun-buffer-val] */
2671                 cfs_hash_del(lock->l_export->exp_lock_hash,
2672                              &lock->l_remote_handle, &lock->l_exp_hash);
2673         }
2674
2675         list_add_tail(&lock->l_rk_ast, rpc_list);
2676         LDLM_LOCK_GET(lock);
2677
2678         unlock_res_and_lock(lock);
2679         return 0;
2680 }
2681
2682 void ldlm_revoke_export_locks(struct obd_export *exp)
2683 {
2684         LIST_HEAD(rpc_list);
2685
2686         ENTRY;
2687
2688         cfs_hash_for_each_nolock(exp->exp_lock_hash,
2689                                  ldlm_revoke_lock_cb, &rpc_list, 0);
2690         ldlm_run_ast_work(exp->exp_obd->obd_namespace, &rpc_list,
2691                           LDLM_WORK_REVOKE_AST);
2692
2693         EXIT;
2694 }
2695 EXPORT_SYMBOL(ldlm_revoke_export_locks);
2696 #endif /* HAVE_SERVER_SUPPORT */
2697
2698 static int ldlm_bl_get_work(struct ldlm_bl_pool *blp,
2699                             struct ldlm_bl_work_item **p_blwi,
2700                             struct obd_export **p_exp)
2701 {
2702         struct ldlm_bl_work_item *blwi = NULL;
2703         static unsigned int num_bl;
2704         static unsigned int num_stale;
2705         int num_th = atomic_read(&blp->blp_num_threads);
2706
2707         *p_exp = obd_stale_export_get();
2708
2709         spin_lock(&blp->blp_lock);
2710         if (*p_exp != NULL) {
2711                 if (num_th == 1 || ++num_stale < num_th) {
2712                         spin_unlock(&blp->blp_lock);
2713                         return 1;
2714                 }
2715                 num_stale = 0;
2716         }
2717
2718         /* process a request from the blp_list at least every blp_num_threads */
2719         if (!list_empty(&blp->blp_list) &&
2720             (list_empty(&blp->blp_prio_list) || num_bl == 0))
2721                 blwi = list_entry(blp->blp_list.next,
2722                                   struct ldlm_bl_work_item, blwi_entry);
2723         else
2724                 if (!list_empty(&blp->blp_prio_list))
2725                         blwi = list_entry(blp->blp_prio_list.next,
2726                                           struct ldlm_bl_work_item,
2727                                           blwi_entry);
2728
2729         if (blwi) {
2730                 if (++num_bl >= num_th)
2731                         num_bl = 0;
2732                 list_del(&blwi->blwi_entry);
2733         }
2734         spin_unlock(&blp->blp_lock);
2735         *p_blwi = blwi;
2736
2737         if (*p_exp != NULL && *p_blwi != NULL) {
2738                 obd_stale_export_put(*p_exp);
2739                 *p_exp = NULL;
2740         }
2741
2742         return (*p_blwi != NULL || *p_exp != NULL) ? 1 : 0;
2743 }
2744
2745 /* This only contains temporary data until the thread starts */
2746 struct ldlm_bl_thread_data {
2747         struct ldlm_bl_pool     *bltd_blp;
2748         struct completion       bltd_comp;
2749         int                     bltd_num;
2750 };
2751
2752 static int ldlm_bl_thread_main(void *arg);
2753
2754 static int ldlm_bl_thread_start(struct ldlm_bl_pool *blp, bool check_busy)
2755 {
2756         struct ldlm_bl_thread_data bltd = { .bltd_blp = blp };
2757         struct task_struct *task;
2758
2759         init_completion(&bltd.bltd_comp);
2760
2761         bltd.bltd_num = atomic_inc_return(&blp->blp_num_threads);
2762         if (bltd.bltd_num >= blp->blp_max_threads) {
2763                 atomic_dec(&blp->blp_num_threads);
2764                 return 0;
2765         }
2766
2767         LASSERTF(bltd.bltd_num > 0, "thread num:%d\n", bltd.bltd_num);
2768         if (check_busy &&
2769             atomic_read(&blp->blp_busy_threads) < (bltd.bltd_num - 1)) {
2770                 atomic_dec(&blp->blp_num_threads);
2771                 return 0;
2772         }
2773
2774         task = kthread_run(ldlm_bl_thread_main, &bltd, "ldlm_bl_%02d",
2775                            bltd.bltd_num);
2776         if (IS_ERR(task)) {
2777                 CERROR("cannot start LDLM thread ldlm_bl_%02d: rc %ld\n",
2778                        bltd.bltd_num, PTR_ERR(task));
2779                 atomic_dec(&blp->blp_num_threads);
2780                 return PTR_ERR(task);
2781         }
2782         wait_for_completion(&bltd.bltd_comp);
2783
2784         return 0;
2785 }
2786
2787 /* Not fatal if racy and have a few too many threads */
2788 static int ldlm_bl_thread_need_create(struct ldlm_bl_pool *blp,
2789                                       struct ldlm_bl_work_item *blwi)
2790 {
2791         if (atomic_read(&blp->blp_num_threads) >= blp->blp_max_threads)
2792                 return 0;
2793
2794         if (atomic_read(&blp->blp_busy_threads) <
2795             atomic_read(&blp->blp_num_threads))
2796                 return 0;
2797
2798         if (blwi != NULL && (blwi->blwi_ns == NULL ||
2799                              blwi->blwi_mem_pressure))
2800                 return 0;
2801
2802         return 1;
2803 }
2804
2805 static int ldlm_bl_thread_blwi(struct ldlm_bl_pool *blp,
2806                                struct ldlm_bl_work_item *blwi)
2807 {
2808         /* '1' for consistency with code that checks !mpflag to restore */
2809         unsigned int mpflags = 1;
2810
2811         ENTRY;
2812
2813         if (blwi->blwi_ns == NULL)
2814                 /* added by ldlm_cleanup() */
2815                 RETURN(LDLM_ITER_STOP);
2816
2817         if (blwi->blwi_mem_pressure)
2818                 mpflags = memalloc_noreclaim_save();
2819
2820         OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_PAUSE_CANCEL2, 4);
2821
2822         if (blwi->blwi_count) {
2823                 int count;
2824                 /*
2825                  * The special case when we cancel locks in lru
2826                  * asynchronously, we pass the list of locks here.
2827                  * Thus locks are marked LDLM_FL_CANCELING, but NOT
2828                  * canceled locally yet.
2829                  */
2830                 count = ldlm_cli_cancel_list_local(&blwi->blwi_head,
2831                                                    blwi->blwi_count,
2832                                                    LCF_BL_AST);
2833                 ldlm_cli_cancel_list(&blwi->blwi_head, count, NULL,
2834                                      blwi->blwi_flags);
2835         } else {
2836                 ldlm_handle_bl_callback(blwi->blwi_ns, &blwi->blwi_ld,
2837                                         blwi->blwi_lock);
2838         }
2839         if (blwi->blwi_mem_pressure)
2840                 memalloc_noreclaim_restore(mpflags);
2841
2842         if (blwi->blwi_flags & LCF_ASYNC)
2843                 OBD_FREE(blwi, sizeof(*blwi));
2844         else
2845                 complete(&blwi->blwi_comp);
2846
2847         RETURN(0);
2848 }
2849
2850 /**
2851  * Cancel stale locks on export. Cancel blocked locks first.
2852  * If the given export has blocked locks, the next in the list may have
2853  * them too, thus cancel not blocked locks only if the current export has
2854  * no blocked locks.
2855  **/
2856 static int ldlm_bl_thread_exports(struct ldlm_bl_pool *blp,
2857                                   struct obd_export *exp)
2858 {
2859         int num;
2860
2861         ENTRY;
2862
2863         OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_BL_EVICT, 4);
2864
2865         num = ldlm_export_cancel_blocked_locks(exp);
2866         if (num == 0)
2867                 ldlm_export_cancel_locks(exp);
2868
2869         obd_stale_export_put(exp);
2870
2871         RETURN(0);
2872 }
2873
2874
2875 /**
2876  * Main blocking requests processing thread.
2877  *
2878  * Callers put locks into its queue by calling ldlm_bl_to_thread.
2879  * This thread in the end ends up doing actual call to ->l_blocking_ast
2880  * for queued locks.
2881  */
2882 static int ldlm_bl_thread_main(void *arg)
2883 {
2884         struct lu_env *env;
2885         struct ldlm_bl_pool *blp;
2886         struct ldlm_bl_thread_data *bltd = arg;
2887         int rc;
2888
2889         ENTRY;
2890
2891         OBD_ALLOC_PTR(env);
2892         if (!env)
2893                 RETURN(-ENOMEM);
2894         rc = lu_env_init(env, LCT_DT_THREAD);
2895         if (rc)
2896                 GOTO(out_env, rc);
2897         rc = lu_env_add(env);
2898         if (rc)
2899                 GOTO(out_env_fini, rc);
2900
2901         blp = bltd->bltd_blp;
2902
2903         complete(&bltd->bltd_comp);
2904         /* cannot use bltd after this, it is only on caller's stack */
2905
2906         while (1) {
2907                 struct ldlm_bl_work_item *blwi = NULL;
2908                 struct obd_export *exp = NULL;
2909                 int rc;
2910
2911                 rc = ldlm_bl_get_work(blp, &blwi, &exp);
2912
2913                 if (rc == 0)
2914                         wait_event_idle_exclusive(blp->blp_waitq,
2915                                                   ldlm_bl_get_work(blp, &blwi,
2916                                                                    &exp));
2917                 atomic_inc(&blp->blp_busy_threads);
2918
2919                 if (ldlm_bl_thread_need_create(blp, blwi))
2920                         /* discard the return value, we tried */
2921                         ldlm_bl_thread_start(blp, true);
2922
2923                 if (exp)
2924                         rc = ldlm_bl_thread_exports(blp, exp);
2925                 else if (blwi)
2926                         rc = ldlm_bl_thread_blwi(blp, blwi);
2927
2928                 atomic_dec(&blp->blp_busy_threads);
2929
2930                 if (rc == LDLM_ITER_STOP)
2931                         break;
2932
2933                 /*
2934                  * If there are many namespaces, we will not sleep waiting for
2935                  * work, and must do a cond_resched to avoid holding the CPU
2936                  * for too long
2937                  */
2938                 cond_resched();
2939         }
2940
2941         atomic_dec(&blp->blp_num_threads);
2942         complete(&blp->blp_comp);
2943
2944         lu_env_remove(env);
2945 out_env_fini:
2946         lu_env_fini(env);
2947 out_env:
2948         OBD_FREE_PTR(env);
2949         RETURN(rc);
2950 }
2951
2952
2953 static int ldlm_setup(void);
2954 static int ldlm_cleanup(void);
2955
2956 int ldlm_get_ref(void)
2957 {
2958         int rc = 0;
2959
2960         ENTRY;
2961         mutex_lock(&ldlm_ref_mutex);
2962         if (++ldlm_refcount == 1) {
2963                 rc = ldlm_setup();
2964                 if (rc)
2965                         ldlm_refcount--;
2966         }
2967         mutex_unlock(&ldlm_ref_mutex);
2968
2969         RETURN(rc);
2970 }
2971
2972 void ldlm_put_ref(void)
2973 {
2974         ENTRY;
2975         mutex_lock(&ldlm_ref_mutex);
2976         if (ldlm_refcount == 1) {
2977                 int rc = ldlm_cleanup();
2978
2979                 if (rc)
2980                         CERROR("ldlm_cleanup failed: %d\n", rc);
2981                 else
2982                         ldlm_refcount--;
2983         } else {
2984                 ldlm_refcount--;
2985         }
2986         mutex_unlock(&ldlm_ref_mutex);
2987
2988         EXIT;
2989 }
2990
2991 /*
2992  * Export handle<->lock hash operations.
2993  */
2994 static unsigned
2995 ldlm_export_lock_hash(struct cfs_hash *hs, const void *key, unsigned int mask)
2996 {
2997         return cfs_hash_u64_hash(((struct lustre_handle *)key)->cookie, mask);
2998 }
2999
3000 static void *
3001 ldlm_export_lock_key(struct hlist_node *hnode)
3002 {
3003         struct ldlm_lock *lock;
3004
3005         lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
3006         return &lock->l_remote_handle;
3007 }
3008
3009 static void
3010 ldlm_export_lock_keycpy(struct hlist_node *hnode, void *key)
3011 {
3012         struct ldlm_lock     *lock;
3013
3014         lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
3015         lock->l_remote_handle = *(struct lustre_handle *)key;
3016 }
3017
3018 static int
3019 ldlm_export_lock_keycmp(const void *key, struct hlist_node *hnode)
3020 {
3021         return lustre_handle_equal(ldlm_export_lock_key(hnode), key);
3022 }
3023
3024 static void *
3025 ldlm_export_lock_object(struct hlist_node *hnode)
3026 {
3027         return hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
3028 }
3029
3030 static void
3031 ldlm_export_lock_get(struct cfs_hash *hs, struct hlist_node *hnode)
3032 {
3033         struct ldlm_lock *lock;
3034
3035         lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
3036         LDLM_LOCK_GET(lock);
3037 }
3038
3039 static void
3040 ldlm_export_lock_put(struct cfs_hash *hs, struct hlist_node *hnode)
3041 {
3042         struct ldlm_lock *lock;
3043
3044         lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
3045         LDLM_LOCK_RELEASE(lock);
3046 }
3047
3048 static struct cfs_hash_ops ldlm_export_lock_ops = {
3049         .hs_hash        = ldlm_export_lock_hash,
3050         .hs_key         = ldlm_export_lock_key,
3051         .hs_keycmp      = ldlm_export_lock_keycmp,
3052         .hs_keycpy      = ldlm_export_lock_keycpy,
3053         .hs_object      = ldlm_export_lock_object,
3054         .hs_get         = ldlm_export_lock_get,
3055         .hs_put         = ldlm_export_lock_put,
3056         .hs_put_locked  = ldlm_export_lock_put,
3057 };
3058
3059 int ldlm_init_export(struct obd_export *exp)
3060 {
3061         int rc;
3062
3063         ENTRY;
3064
3065         exp->exp_lock_hash =
3066                 cfs_hash_create(obd_uuid2str(&exp->exp_client_uuid),
3067                                 HASH_EXP_LOCK_CUR_BITS,
3068                                 HASH_EXP_LOCK_MAX_BITS,
3069                                 HASH_EXP_LOCK_BKT_BITS, 0,
3070                                 CFS_HASH_MIN_THETA, CFS_HASH_MAX_THETA,
3071                                 &ldlm_export_lock_ops,
3072                                 CFS_HASH_DEFAULT | CFS_HASH_REHASH_KEY |
3073                                 CFS_HASH_NBLK_CHANGE);
3074
3075         if (!exp->exp_lock_hash)
3076                 RETURN(-ENOMEM);
3077
3078         rc = ldlm_init_flock_export(exp);
3079         if (rc)
3080                 GOTO(err, rc);
3081
3082         RETURN(0);
3083 err:
3084         ldlm_destroy_export(exp);
3085         RETURN(rc);
3086 }
3087 EXPORT_SYMBOL(ldlm_init_export);
3088
3089 void ldlm_destroy_export(struct obd_export *exp)
3090 {
3091         ENTRY;
3092         cfs_hash_putref(exp->exp_lock_hash);
3093         exp->exp_lock_hash = NULL;
3094
3095         ldlm_destroy_flock_export(exp);
3096         EXIT;
3097 }
3098 EXPORT_SYMBOL(ldlm_destroy_export);
3099
3100 static ssize_t cancel_unused_locks_before_replay_show(struct kobject *kobj,
3101                                                       struct attribute *attr,
3102                                                       char *buf)
3103 {
3104         return sprintf(buf, "%d\n", ldlm_cancel_unused_locks_before_replay);
3105 }
3106
3107 static ssize_t cancel_unused_locks_before_replay_store(struct kobject *kobj,
3108                                                        struct attribute *attr,
3109                                                        const char *buffer,
3110                                                        size_t count)
3111 {
3112         int rc;
3113         unsigned long val;
3114
3115         rc = kstrtoul(buffer, 10, &val);
3116         if (rc)
3117                 return rc;
3118
3119         ldlm_cancel_unused_locks_before_replay = val;
3120
3121         return count;
3122 }
3123 LUSTRE_RW_ATTR(cancel_unused_locks_before_replay);
3124
3125 static struct attribute *ldlm_attrs[] = {
3126         &lustre_attr_cancel_unused_locks_before_replay.attr,
3127         NULL,
3128 };
3129
3130 static struct attribute_group ldlm_attr_group = {
3131         .attrs = ldlm_attrs,
3132 };
3133
3134 static int ldlm_setup(void)
3135 {
3136         static struct ptlrpc_service_conf       conf;
3137         struct ldlm_bl_pool                    *blp = NULL;
3138 #ifdef HAVE_SERVER_SUPPORT
3139         struct task_struct *task;
3140 #endif /* HAVE_SERVER_SUPPORT */
3141         int i;
3142         int rc = 0;
3143
3144         ENTRY;
3145
3146         if (ldlm_state != NULL)
3147                 RETURN(-EALREADY);
3148
3149         OBD_ALLOC(ldlm_state, sizeof(*ldlm_state));
3150         if (ldlm_state == NULL)
3151                 RETURN(-ENOMEM);
3152
3153         ldlm_kobj = kobject_create_and_add("ldlm", &lustre_kset->kobj);
3154         if (!ldlm_kobj)
3155                 GOTO(out, -ENOMEM);
3156
3157         rc = sysfs_create_group(ldlm_kobj, &ldlm_attr_group);
3158         if (rc)
3159                 GOTO(out, rc);
3160
3161         ldlm_ns_kset = kset_create_and_add("namespaces", NULL, ldlm_kobj);
3162         if (!ldlm_ns_kset)
3163                 GOTO(out, -ENOMEM);
3164
3165         ldlm_svc_kset = kset_create_and_add("services", NULL, ldlm_kobj);
3166         if (!ldlm_svc_kset)
3167                 GOTO(out, -ENOMEM);
3168
3169         rc = ldlm_debugfs_setup();
3170         if (rc != 0)
3171                 GOTO(out, rc);
3172
3173         memset(&conf, 0, sizeof(conf));
3174         conf = (typeof(conf)) {
3175                 .psc_name               = "ldlm_cbd",
3176                 .psc_watchdog_factor    = 2,
3177                 .psc_buf                = {
3178                         .bc_nbufs               = LDLM_CLIENT_NBUFS,
3179                         .bc_buf_size            = LDLM_BUFSIZE,
3180                         .bc_req_max_size        = LDLM_MAXREQSIZE,
3181                         .bc_rep_max_size        = LDLM_MAXREPSIZE,
3182                         .bc_req_portal          = LDLM_CB_REQUEST_PORTAL,
3183                         .bc_rep_portal          = LDLM_CB_REPLY_PORTAL,
3184                 },
3185                 .psc_thr                = {
3186                         .tc_thr_name            = "ldlm_cb",
3187                         .tc_thr_factor          = LDLM_THR_FACTOR,
3188                         .tc_nthrs_init          = LDLM_NTHRS_INIT,
3189                         .tc_nthrs_base          = LDLM_NTHRS_BASE,
3190                         .tc_nthrs_max           = LDLM_NTHRS_MAX,
3191                         .tc_nthrs_user          = ldlm_num_threads,
3192                         .tc_cpu_bind            = ldlm_cpu_bind,
3193                         .tc_ctx_tags            = LCT_MD_THREAD | LCT_DT_THREAD,
3194                 },
3195                 .psc_cpt                = {
3196                         .cc_pattern             = ldlm_cpts,
3197                         .cc_affinity            = true,
3198                 },
3199                 .psc_ops                = {
3200                         .so_req_handler         = ldlm_callback_handler,
3201                 },
3202         };
3203         ldlm_state->ldlm_cb_service = \
3204                         ptlrpc_register_service(&conf, ldlm_svc_kset,
3205                                                 ldlm_svc_debugfs_dir);
3206         if (IS_ERR(ldlm_state->ldlm_cb_service)) {
3207                 CERROR("failed to start service\n");
3208                 rc = PTR_ERR(ldlm_state->ldlm_cb_service);
3209                 ldlm_state->ldlm_cb_service = NULL;
3210                 GOTO(out, rc);
3211         }
3212
3213 #ifdef HAVE_SERVER_SUPPORT
3214         memset(&conf, 0, sizeof(conf));
3215         conf = (typeof(conf)) {
3216                 .psc_name               = "ldlm_canceld",
3217                 .psc_watchdog_factor    = 6,
3218                 .psc_buf                = {