Whamcloud - gitweb
LU-14312 ldlm: don't change GROUP lock GID on client
[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_handle[1].cookie = lock->l_handle.h_cookie;
948         body->lock_desc = *desc;
949         body->lock_flags |= ldlm_flags_to_wire(lock->l_flags & LDLM_FL_AST_MASK);
950
951         LDLM_DEBUG(lock, "server preparing blocking AST");
952
953         ptlrpc_request_set_replen(req);
954         ldlm_set_cbpending(lock);
955         if (instant_cancel) {
956                 unlock_res_and_lock(lock);
957                 ldlm_lock_cancel(lock);
958
959                 req->rq_no_resend = 1;
960         } else {
961                 LASSERT(ldlm_is_granted(lock));
962                 ldlm_add_waiting_lock(lock, ldlm_bl_timeout(lock));
963                 unlock_res_and_lock(lock);
964
965                 /* Do not resend after lock callback timeout */
966                 req->rq_delay_limit = ldlm_bl_timeout(lock);
967                 req->rq_resend_cb = ldlm_update_resend;
968         }
969
970         req->rq_send_state = LUSTRE_IMP_FULL;
971         /* ptlrpc_request_alloc_pack already set timeout */
972         if (AT_OFF)
973                 req->rq_timeout = ldlm_get_rq_timeout();
974
975         if (lock->l_export && lock->l_export->exp_nid_stats &&
976             lock->l_export->exp_nid_stats->nid_ldlm_stats)
977                 lprocfs_counter_incr(lock->l_export->exp_nid_stats->nid_ldlm_stats,
978                                      LDLM_BL_CALLBACK - LDLM_FIRST_OPC);
979
980         rc = ldlm_ast_fini(req, arg, lock, instant_cancel);
981
982         RETURN(rc);
983 }
984
985 /**
986  * ->l_completion_ast callback for a remote lock in server namespace.
987  *
988  *  Sends AST to the client notifying it of lock granting.  If initial
989  *  lock response was not sent yet, instead of sending another RPC, just
990  *  mark the lock as granted and client will understand
991  */
992 int ldlm_server_completion_ast(struct ldlm_lock *lock, __u64 flags, void *data)
993 {
994         struct ldlm_cb_set_arg *arg = data;
995         struct ldlm_request *body;
996         struct ptlrpc_request *req;
997         struct ldlm_cb_async_args *ca;
998         int instant_cancel = 0;
999         int rc = 0;
1000         int lvb_len;
1001
1002         ENTRY;
1003
1004         LASSERT(lock != NULL);
1005         LASSERT(data != NULL);
1006
1007         if (OBD_FAIL_PRECHECK(OBD_FAIL_LDLM_SRV_CP_AST)) {
1008                 LDLM_DEBUG(lock, "dropping CP AST");
1009                 RETURN(0);
1010         }
1011
1012         req = ptlrpc_request_alloc(lock->l_export->exp_imp_reverse,
1013                                    &RQF_LDLM_CP_CALLBACK);
1014         if (req == NULL)
1015                 RETURN(-ENOMEM);
1016
1017         /* server namespace, doesn't need lock */
1018         lvb_len = ldlm_lvbo_size(lock);
1019         /*
1020          * LU-3124 & LU-2187: to not return layout in completion AST because
1021          * it may deadlock for LU-2187, or client may not have enough space
1022          * for large layout. The layout will be returned to client with an
1023          * extra RPC to fetch xattr.lov
1024          */
1025         if (ldlm_has_layout(lock))
1026                 lvb_len = 0;
1027
1028         req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_CLIENT, lvb_len);
1029         rc = ptlrpc_request_pack(req, LUSTRE_DLM_VERSION, LDLM_CP_CALLBACK);
1030         if (rc) {
1031                 ptlrpc_request_free(req);
1032                 RETURN(rc);
1033         }
1034
1035         ca = ptlrpc_req_async_args(ca, req);
1036         ca->ca_set_arg = arg;
1037         ca->ca_lock = lock;
1038
1039         req->rq_interpret_reply = ldlm_cb_interpret;
1040         body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1041
1042         body->lock_handle[0] = lock->l_remote_handle;
1043         body->lock_handle[1].cookie = lock->l_handle.h_cookie;
1044         body->lock_flags = ldlm_flags_to_wire(flags);
1045         ldlm_lock2desc(lock, &body->lock_desc);
1046         if (lvb_len > 0) {
1047                 void *lvb = req_capsule_client_get(&req->rq_pill, &RMF_DLM_LVB);
1048                 lvb_len = ldlm_lvbo_fill(lock, lvb, &lvb_len);
1049                 if (lvb_len < 0) {
1050                         /*
1051                          * We still need to send the RPC to wake up the blocked
1052                          * enqueue thread on the client.
1053                          *
1054                          * Consider old client, there is no better way to notify
1055                          * the failure, just zero-sized the LVB, then the client
1056                          * will fail out as "-EPROTO".
1057                          */
1058                         req_capsule_shrink(&req->rq_pill, &RMF_DLM_LVB, 0,
1059                                            RCL_CLIENT);
1060                         instant_cancel = 1;
1061                 } else {
1062                         req_capsule_shrink(&req->rq_pill, &RMF_DLM_LVB, lvb_len,
1063                                            RCL_CLIENT);
1064                 }
1065         }
1066
1067         LDLM_DEBUG(lock, "server preparing completion AST");
1068
1069         ptlrpc_request_set_replen(req);
1070
1071         req->rq_send_state = LUSTRE_IMP_FULL;
1072         /* ptlrpc_request_pack already set timeout */
1073         if (AT_OFF)
1074                 req->rq_timeout = ldlm_get_rq_timeout();
1075
1076         /* We only send real blocking ASTs after the lock is granted */
1077         lock_res_and_lock(lock);
1078         if (ldlm_is_ast_sent(lock)) {
1079                 body->lock_flags |= ldlm_flags_to_wire(LDLM_FL_AST_SENT);
1080                 /* Copy AST flags like LDLM_FL_DISCARD_DATA. */
1081                 body->lock_flags |= ldlm_flags_to_wire(lock->l_flags &
1082                                                        LDLM_FL_AST_MASK);
1083
1084                 /*
1085                  * We might get here prior to ldlm_handle_enqueue setting
1086                  * LDLM_FL_CANCEL_ON_BLOCK flag. Then we will put this lock
1087                  * into waiting list, but this is safe and similar code in
1088                  * ldlm_handle_enqueue will call ldlm_lock_cancel() still,
1089                  * that would not only cancel the lock, but will also remove
1090                  * it from waiting list
1091                  */
1092                 if (ldlm_is_cancel_on_block(lock)) {
1093                         unlock_res_and_lock(lock);
1094                         ldlm_lock_cancel(lock);
1095
1096                         instant_cancel = 1;
1097                         req->rq_no_resend = 1;
1098
1099                         lock_res_and_lock(lock);
1100                 } else {
1101                         /* start the lock-timeout clock */
1102                         ldlm_add_waiting_lock(lock, ldlm_bl_timeout(lock));
1103                         /* Do not resend after lock callback timeout */
1104                         req->rq_delay_limit = ldlm_bl_timeout(lock);
1105                         req->rq_resend_cb = ldlm_update_resend;
1106                 }
1107         }
1108         unlock_res_and_lock(lock);
1109
1110         if (lock->l_export && lock->l_export->exp_nid_stats &&
1111             lock->l_export->exp_nid_stats->nid_ldlm_stats)
1112                 lprocfs_counter_incr(lock->l_export->exp_nid_stats->nid_ldlm_stats,
1113                                      LDLM_CP_CALLBACK - LDLM_FIRST_OPC);
1114
1115         rc = ldlm_ast_fini(req, arg, lock, instant_cancel);
1116
1117         RETURN(lvb_len < 0 ? lvb_len : rc);
1118 }
1119
1120 /**
1121  * Server side ->l_glimpse_ast handler for client locks.
1122  *
1123  * Sends glimpse AST to the client and waits for reply. Then updates
1124  * lvbo with the result.
1125  */
1126 int ldlm_server_glimpse_ast(struct ldlm_lock *lock, void *data)
1127 {
1128         struct ldlm_cb_set_arg *arg = data;
1129         struct ldlm_request *body;
1130         struct ptlrpc_request *req;
1131         struct ldlm_cb_async_args *ca;
1132         int rc;
1133         struct req_format *req_fmt;
1134
1135         ENTRY;
1136
1137         LASSERT(lock != NULL);
1138
1139         if (arg->gl_desc != NULL)
1140                 /* There is a glimpse descriptor to pack */
1141                 req_fmt = &RQF_LDLM_GL_CALLBACK_DESC;
1142         else
1143                 req_fmt = &RQF_LDLM_GL_CALLBACK;
1144
1145         req = ptlrpc_request_alloc_pack(lock->l_export->exp_imp_reverse,
1146                                         req_fmt, LUSTRE_DLM_VERSION,
1147                                         LDLM_GL_CALLBACK);
1148
1149         if (req == NULL)
1150                 RETURN(-ENOMEM);
1151
1152         if (arg->gl_desc != NULL) {
1153                 /* copy the GL descriptor */
1154                 union ldlm_gl_desc      *desc;
1155
1156                 desc = req_capsule_client_get(&req->rq_pill, &RMF_DLM_GL_DESC);
1157                 *desc = *arg->gl_desc;
1158         }
1159
1160         body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1161         body->lock_handle[0] = lock->l_remote_handle;
1162         ldlm_lock2desc(lock, &body->lock_desc);
1163
1164         ca = ptlrpc_req_async_args(ca, req);
1165         ca->ca_set_arg = arg;
1166         ca->ca_lock = lock;
1167
1168         /* server namespace, doesn't need lock */
1169         req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
1170                              ldlm_lvbo_size(lock));
1171         ptlrpc_request_set_replen(req);
1172
1173         req->rq_send_state = LUSTRE_IMP_FULL;
1174         /* ptlrpc_request_alloc_pack already set timeout */
1175         if (AT_OFF)
1176                 req->rq_timeout = ldlm_get_rq_timeout();
1177
1178         req->rq_interpret_reply = ldlm_cb_interpret;
1179
1180         if (lock->l_export && lock->l_export->exp_nid_stats) {
1181                 struct nid_stat *nid_stats = lock->l_export->exp_nid_stats;
1182
1183                 lprocfs_counter_incr(nid_stats->nid_ldlm_stats,
1184                                      LDLM_GL_CALLBACK - LDLM_FIRST_OPC);
1185         }
1186
1187         rc = ldlm_ast_fini(req, arg, lock, 0);
1188
1189         RETURN(rc);
1190 }
1191 EXPORT_SYMBOL(ldlm_server_glimpse_ast);
1192
1193 int ldlm_glimpse_locks(struct ldlm_resource *res,
1194                        struct list_head *gl_work_list)
1195 {
1196         int rc;
1197
1198         ENTRY;
1199
1200         rc = ldlm_run_ast_work(ldlm_res_to_ns(res), gl_work_list,
1201                                LDLM_WORK_GL_AST);
1202         if (rc == -ERESTART)
1203                 ldlm_reprocess_all(res, NULL);
1204
1205         RETURN(rc);
1206 }
1207 EXPORT_SYMBOL(ldlm_glimpse_locks);
1208
1209 /* return LDLM lock associated with a lock callback request */
1210 struct ldlm_lock *ldlm_request_lock(struct ptlrpc_request *req)
1211 {
1212         struct ldlm_cb_async_args *ca;
1213         struct ldlm_lock *lock;
1214
1215         ENTRY;
1216
1217         ca = ptlrpc_req_async_args(ca, req);
1218         lock = ca->ca_lock;
1219         if (lock == NULL)
1220                 RETURN(ERR_PTR(-EFAULT));
1221
1222         RETURN(lock);
1223 }
1224 EXPORT_SYMBOL(ldlm_request_lock);
1225
1226 /**
1227  * Main server-side entry point into LDLM for enqueue. This is called by ptlrpc
1228  * service threads to carry out client lock enqueueing requests.
1229  */
1230 int ldlm_handle_enqueue0(struct ldlm_namespace *ns,
1231                          struct ptlrpc_request *req,
1232                          const struct ldlm_request *dlm_req,
1233                          const struct ldlm_callback_suite *cbs)
1234 {
1235         struct ldlm_reply *dlm_rep;
1236         __u64 flags;
1237         enum ldlm_error err = ELDLM_OK;
1238         struct ldlm_lock *lock = NULL;
1239         void *cookie = NULL;
1240         int rc = 0;
1241         struct ldlm_resource *res = NULL;
1242         const struct lu_env *env = req->rq_svc_thread->t_env;
1243
1244         ENTRY;
1245
1246         LDLM_DEBUG_NOLOCK("server-side enqueue handler START");
1247
1248         ldlm_request_cancel(req, dlm_req, LDLM_ENQUEUE_CANCEL_OFF, LATF_SKIP);
1249         flags = ldlm_flags_from_wire(dlm_req->lock_flags);
1250
1251         LASSERT(req->rq_export);
1252
1253         /* for intent enqueue the stat will be updated inside intent policy */
1254         if (ptlrpc_req2svc(req)->srv_stats != NULL &&
1255             !(dlm_req->lock_flags & LDLM_FL_HAS_INTENT))
1256                 ldlm_svc_get_eopc(dlm_req, ptlrpc_req2svc(req)->srv_stats);
1257
1258         if (req->rq_export && req->rq_export->exp_nid_stats &&
1259             req->rq_export->exp_nid_stats->nid_ldlm_stats)
1260                 lprocfs_counter_incr(req->rq_export->exp_nid_stats->nid_ldlm_stats,
1261                                      LDLM_ENQUEUE - LDLM_FIRST_OPC);
1262
1263         if (unlikely(dlm_req->lock_desc.l_resource.lr_type < LDLM_MIN_TYPE ||
1264                      dlm_req->lock_desc.l_resource.lr_type >= LDLM_MAX_TYPE)) {
1265                 DEBUG_REQ(D_ERROR, req, "invalid lock request type %d",
1266                           dlm_req->lock_desc.l_resource.lr_type);
1267                 GOTO(out, rc = -EFAULT);
1268         }
1269
1270         if (unlikely(dlm_req->lock_desc.l_req_mode <= LCK_MINMODE ||
1271                      dlm_req->lock_desc.l_req_mode >= LCK_MAXMODE ||
1272                      dlm_req->lock_desc.l_req_mode &
1273                      (dlm_req->lock_desc.l_req_mode-1))) {
1274                 DEBUG_REQ(D_ERROR, req, "invalid lock request mode %d",
1275                           dlm_req->lock_desc.l_req_mode);
1276                 GOTO(out, rc = -EFAULT);
1277         }
1278
1279         if (unlikely((flags & LDLM_FL_REPLAY) ||
1280                      (lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT))) {
1281                 /* Find an existing lock in the per-export lock hash */
1282                 /*
1283                  * In the function below, .hs_keycmp resolves to
1284                  * ldlm_export_lock_keycmp()
1285                  */
1286                 /* coverity[overrun-buffer-val] */
1287                 lock = cfs_hash_lookup(req->rq_export->exp_lock_hash,
1288                                        (void *)&dlm_req->lock_handle[0]);
1289                 if (lock != NULL) {
1290                         DEBUG_REQ(D_DLMTRACE, req,
1291                                   "found existing lock cookie %#llx",
1292                                   lock->l_handle.h_cookie);
1293                         flags |= LDLM_FL_RESENT;
1294                         GOTO(existing_lock, rc = 0);
1295                 }
1296         } else {
1297                 if (ldlm_reclaim_full()) {
1298                         DEBUG_REQ(D_DLMTRACE, req,
1299                                   "Too many granted locks, reject current enqueue request and let the client retry later");
1300                         GOTO(out, rc = -EINPROGRESS);
1301                 }
1302         }
1303
1304         /* The lock's callback data might be set in the policy function */
1305         lock = ldlm_lock_create(ns, &dlm_req->lock_desc.l_resource.lr_name,
1306                                 dlm_req->lock_desc.l_resource.lr_type,
1307                                 dlm_req->lock_desc.l_req_mode,
1308                                 cbs, NULL, 0, LVB_T_NONE);
1309         if (IS_ERR(lock)) {
1310                 rc = PTR_ERR(lock);
1311                 lock = NULL;
1312                 GOTO(out, rc);
1313         }
1314
1315         lock->l_remote_handle = dlm_req->lock_handle[0];
1316         LDLM_DEBUG(lock, "server-side enqueue handler, new lock created");
1317
1318         /*
1319          * Initialize resource lvb but not for a lock being replayed since
1320          * Client already got lvb sent in this case.
1321          * This must occur early since some policy methods assume resource
1322          * lvb is available (lr_lvb_data != NULL).
1323          */
1324         res = lock->l_resource;
1325         if (!(flags & LDLM_FL_REPLAY)) {
1326                 /* non-replayed lock, delayed lvb init may need to be done */
1327                 rc = ldlm_lvbo_init(res);
1328                 if (rc < 0) {
1329                         LDLM_DEBUG(lock, "delayed lvb init failed (rc %d)", rc);
1330                         GOTO(out, rc);
1331                 }
1332         }
1333
1334         OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_ENQUEUE_BLOCKED, obd_timeout * 2);
1335         /*
1336          * Don't enqueue a lock onto the export if it is been disonnected
1337          * due to eviction (b=3822) or server umount (b=24324).
1338          * Cancel it now instead.
1339          */
1340         if (req->rq_export->exp_disconnected) {
1341                 LDLM_ERROR(lock, "lock on disconnected export %p",
1342                            req->rq_export);
1343                 GOTO(out, rc = -ENOTCONN);
1344         }
1345
1346         lock->l_export = class_export_lock_get(req->rq_export, lock);
1347         if (lock->l_export->exp_lock_hash)
1348                 cfs_hash_add(lock->l_export->exp_lock_hash,
1349                              &lock->l_remote_handle,
1350                              &lock->l_exp_hash);
1351
1352         /*
1353          * Inherit the enqueue flags before the operation, because we do not
1354          * keep the res lock on return and next operations (BL AST) may proceed
1355          * without them.
1356          */
1357         lock->l_flags |= ldlm_flags_from_wire(dlm_req->lock_flags &
1358                                               LDLM_FL_INHERIT_MASK);
1359
1360         ldlm_convert_policy_to_local(req->rq_export,
1361                                      dlm_req->lock_desc.l_resource.lr_type,
1362                                      &dlm_req->lock_desc.l_policy_data,
1363                                      &lock->l_policy_data);
1364         if (dlm_req->lock_desc.l_resource.lr_type == LDLM_EXTENT) {
1365                 lock->l_req_extent = lock->l_policy_data.l_extent;
1366         } else if (dlm_req->lock_desc.l_resource.lr_type == LDLM_IBITS) {
1367                 lock->l_policy_data.l_inodebits.try_bits =
1368                         dlm_req->lock_desc.l_policy_data.l_inodebits.try_bits;
1369                 lock->l_policy_data.l_inodebits.li_gid =
1370                         dlm_req->lock_desc.l_policy_data.l_inodebits.li_gid;
1371         }
1372
1373 existing_lock:
1374         cookie = req;
1375         if (!(flags & LDLM_FL_HAS_INTENT)) {
1376                 /* based on the assumption that lvb size never changes during
1377                  * resource life time otherwise it need resource->lr_lock's
1378                  * protection */
1379                 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB,
1380                                      RCL_SERVER, ldlm_lvbo_size(lock));
1381
1382                 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_ENQUEUE_EXTENT_ERR))
1383                         GOTO(out, rc = -ENOMEM);
1384
1385                 rc = req_capsule_server_pack(&req->rq_pill);
1386                 if (rc)
1387                         GOTO(out, rc);
1388         }
1389
1390         err = ldlm_lock_enqueue(env, ns, &lock, cookie, &flags);
1391         if (err) {
1392                 if ((int)err < 0)
1393                         rc = (int)err;
1394                 GOTO(out, err);
1395         }
1396
1397         dlm_rep = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
1398
1399         ldlm_lock2desc(lock, &dlm_rep->lock_desc);
1400         ldlm_lock2handle(lock, &dlm_rep->lock_handle);
1401
1402         if (lock && lock->l_resource->lr_type == LDLM_EXTENT)
1403                 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_BL_EVICT, 6);
1404
1405         /*
1406          * We never send a blocking AST until the lock is granted, but
1407          * we can tell it right now
1408          */
1409         lock_res_and_lock(lock);
1410
1411         /*
1412          * Now take into account flags to be inherited from original lock
1413          * request both in reply to client and in our own lock flags.
1414          */
1415         dlm_rep->lock_flags = ldlm_flags_to_wire(flags);
1416         lock->l_flags |= flags & LDLM_FL_INHERIT_MASK;
1417
1418         /*
1419          * Don't move a pending lock onto the export if it has already been
1420          * disconnected due to eviction (b=5683) or server umount (b=24324).
1421          * Cancel it now instead.
1422          */
1423         if (unlikely(req->rq_export->exp_disconnected ||
1424                      OBD_FAIL_CHECK(OBD_FAIL_LDLM_ENQUEUE_OLD_EXPORT))) {
1425                 LDLM_ERROR(lock, "lock on destroyed export %p", req->rq_export);
1426                 rc = -ENOTCONN;
1427         } else if (ldlm_is_ast_sent(lock)) {
1428                 /* fill lock desc for possible lock convert */
1429                 if (lock->l_blocking_lock &&
1430                     lock->l_resource->lr_type == LDLM_IBITS) {
1431                         struct ldlm_lock *bl_lock = lock->l_blocking_lock;
1432                         struct ldlm_lock_desc *rep_desc = &dlm_rep->lock_desc;
1433
1434                         LDLM_DEBUG(lock,
1435                                    "save blocking bits %llx in granted lock",
1436                                    bl_lock->l_policy_data.l_inodebits.bits);
1437                         /*
1438                          * If lock is blocked then save blocking ibits
1439                          * in returned lock policy for the possible lock
1440                          * convert on a client.
1441                          */
1442                         rep_desc->l_policy_data.l_inodebits.cancel_bits =
1443                                 bl_lock->l_policy_data.l_inodebits.bits;
1444                 }
1445                 dlm_rep->lock_flags |= ldlm_flags_to_wire(LDLM_FL_AST_SENT);
1446                 if (ldlm_is_granted(lock)) {
1447                         /*
1448                          * Only cancel lock if it was granted, because it would
1449                          * be destroyed immediately and would never be granted
1450                          * in the future, causing timeouts on client.  Not
1451                          * granted lock will be cancelled immediately after
1452                          * sending completion AST.
1453                          */
1454                         if (ldlm_is_cancel_on_block(lock)) {
1455                                 unlock_res_and_lock(lock);
1456                                 ldlm_lock_cancel(lock);
1457                                 lock_res_and_lock(lock);
1458                         } else {
1459                                 ldlm_add_waiting_lock(lock,
1460                                                       ldlm_bl_timeout(lock));
1461                         }
1462                 }
1463         }
1464         unlock_res_and_lock(lock);
1465
1466         EXIT;
1467 out:
1468         req->rq_status = rc ?: err; /* return either error - b=11190 */
1469         if (!req->rq_packed_final) {
1470                 err = lustre_pack_reply(req, 1, NULL, NULL);
1471                 if (rc == 0)
1472                         rc = err;
1473         }
1474
1475         /*
1476          * The LOCK_CHANGED code in ldlm_lock_enqueue depends on this
1477          * ldlm_reprocess_all.  If this moves, revisit that code. -phil
1478          */
1479         if (lock != NULL) {
1480                 LDLM_DEBUG(lock,
1481                            "server-side enqueue handler, sending reply (err=%d, rc=%d)",
1482                            err, rc);
1483
1484                 if (rc == 0 &&
1485                     req_capsule_has_field(&req->rq_pill, &RMF_DLM_LVB,
1486                                           RCL_SERVER) &&
1487                     ldlm_lvbo_size(lock) > 0) {
1488                         void *buf;
1489                         int buflen;
1490
1491 retry:
1492                         buf = req_capsule_server_get(&req->rq_pill,
1493                                                      &RMF_DLM_LVB);
1494                         LASSERTF(buf != NULL, "req %p, lock %p\n", req, lock);
1495                         buflen = req_capsule_get_size(&req->rq_pill,
1496                                         &RMF_DLM_LVB, RCL_SERVER);
1497                         /*
1498                          * non-replayed lock, delayed lvb init may
1499                          * need to be occur now
1500                          */
1501                         if ((buflen > 0) && !(flags & LDLM_FL_REPLAY)) {
1502                                 int rc2;
1503
1504                                 rc2 = ldlm_lvbo_fill(lock, buf, &buflen);
1505                                 if (rc2 >= 0) {
1506                                         req_capsule_shrink(&req->rq_pill,
1507                                                            &RMF_DLM_LVB,
1508                                                            rc2, RCL_SERVER);
1509                                 } else if (rc2 == -ERANGE) {
1510                                         rc2 = req_capsule_server_grow(
1511                                                         &req->rq_pill,
1512                                                         &RMF_DLM_LVB, buflen);
1513                                         if (!rc2) {
1514                                                 goto retry;
1515                                         } else {
1516                                                 /*
1517                                                  * if we can't grow the buffer,
1518                                                  * it's ok to return empty lvb
1519                                                  * to client.
1520                                                  */
1521                                                 req_capsule_shrink(
1522                                                         &req->rq_pill,
1523                                                         &RMF_DLM_LVB, 0,
1524                                                         RCL_SERVER);
1525                                         }
1526                                 } else {
1527                                         rc = rc2;
1528                                 }
1529                         } else if (flags & LDLM_FL_REPLAY) {
1530                                 /* no LVB resend upon replay */
1531                                 if (buflen > 0)
1532                                         req_capsule_shrink(&req->rq_pill,
1533                                                            &RMF_DLM_LVB,
1534                                                            0, RCL_SERVER);
1535                                 else
1536                                         rc = buflen;
1537                         } else {
1538                                 rc = buflen;
1539                         }
1540                 }
1541
1542                 if (rc != 0 && !(flags & LDLM_FL_RESENT)) {
1543                         if (lock->l_export) {
1544                                 ldlm_lock_cancel(lock);
1545                         } else {
1546                                 lock_res_and_lock(lock);
1547                                 ldlm_resource_unlink_lock(lock);
1548                                 ldlm_lock_destroy_nolock(lock);
1549                                 unlock_res_and_lock(lock);
1550
1551                         }
1552                 }
1553
1554                 if (!err && !ldlm_is_cbpending(lock) &&
1555                     dlm_req->lock_desc.l_resource.lr_type != LDLM_FLOCK)
1556                         ldlm_reprocess_all(lock->l_resource, lock);
1557
1558                 LDLM_LOCK_RELEASE(lock);
1559         }
1560
1561         LDLM_DEBUG_NOLOCK("server-side enqueue handler END (lock %p, rc %d)",
1562                           lock, rc);
1563
1564         return rc;
1565 }
1566
1567 /*
1568  * Clear the blocking lock, the race is possible between ldlm_handle_convert0()
1569  * and ldlm_work_bl_ast_lock(), so this is done under lock with check for NULL.
1570  */
1571 void ldlm_clear_blocking_lock(struct ldlm_lock *lock)
1572 {
1573         if (lock->l_blocking_lock) {
1574                 LDLM_LOCK_RELEASE(lock->l_blocking_lock);
1575                 lock->l_blocking_lock = NULL;
1576         }
1577 }
1578
1579 /* A lock can be converted to new ibits or mode and should be considered
1580  * as new lock. Clear all states related to a previous blocking AST
1581  * processing so new conflicts will cause new blocking ASTs.
1582  *
1583  * This is used during lock convert below and lock downgrade to COS mode in
1584  * ldlm_lock_mode_downgrade().
1585  */
1586 void ldlm_clear_blocking_data(struct ldlm_lock *lock)
1587 {
1588         ldlm_clear_ast_sent(lock);
1589         lock->l_bl_ast_run = 0;
1590         ldlm_clear_blocking_lock(lock);
1591 }
1592
1593 /**
1594  * Main LDLM entry point for server code to process lock conversion requests.
1595  */
1596 int ldlm_handle_convert0(struct ptlrpc_request *req,
1597                          const struct ldlm_request *dlm_req)
1598 {
1599         struct obd_export *exp = req->rq_export;
1600         struct ldlm_reply *dlm_rep;
1601         struct ldlm_lock *lock;
1602         __u64 bits;
1603         __u64 new_bits;
1604         int rc;
1605
1606         ENTRY;
1607
1608         if (exp && exp->exp_nid_stats && exp->exp_nid_stats->nid_ldlm_stats)
1609                 lprocfs_counter_incr(exp->exp_nid_stats->nid_ldlm_stats,
1610                                      LDLM_CONVERT - LDLM_FIRST_OPC);
1611
1612         rc = req_capsule_server_pack(&req->rq_pill);
1613         if (rc)
1614                 RETURN(rc);
1615
1616         dlm_rep = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
1617         dlm_rep->lock_flags = dlm_req->lock_flags;
1618
1619         lock = ldlm_handle2lock(&dlm_req->lock_handle[0]);
1620         if (!lock) {
1621                 LDLM_DEBUG_NOLOCK("server lock is canceled already");
1622                 req->rq_status = ELDLM_NO_LOCK_DATA;
1623                 RETURN(0);
1624         }
1625
1626         LDLM_DEBUG(lock, "server-side convert handler START");
1627
1628         lock_res_and_lock(lock);
1629         bits = lock->l_policy_data.l_inodebits.bits;
1630         new_bits = dlm_req->lock_desc.l_policy_data.l_inodebits.bits;
1631
1632         if (ldlm_is_cancel(lock)) {
1633                 LDLM_DEBUG(lock, "convert on canceled lock!");
1634                 unlock_res_and_lock(lock);
1635                 GOTO(out_put, rc = ELDLM_NO_LOCK_DATA);
1636         }
1637
1638         if (dlm_req->lock_desc.l_req_mode != lock->l_granted_mode) {
1639                 LDLM_ERROR(lock, "lock mode differs!");
1640                 unlock_res_and_lock(lock);
1641                 GOTO(out_put, rc = -EPROTO);
1642         }
1643
1644         if (bits == new_bits) {
1645                 /*
1646                  * This can be valid situation if CONVERT RPCs are
1647                  * re-ordered. Just finish silently
1648                  */
1649                 LDLM_DEBUG(lock, "lock is converted already!");
1650                 unlock_res_and_lock(lock);
1651         } else {
1652                 if (ldlm_is_waited(lock))
1653                         ldlm_del_waiting_lock(lock);
1654
1655                 ldlm_clear_cbpending(lock);
1656                 lock->l_policy_data.l_inodebits.cancel_bits = 0;
1657                 ldlm_inodebits_drop(lock, bits & ~new_bits);
1658
1659                 ldlm_clear_blocking_data(lock);
1660                 unlock_res_and_lock(lock);
1661
1662                 ldlm_reprocess_all(lock->l_resource, NULL);
1663         }
1664
1665         dlm_rep->lock_handle = lock->l_remote_handle;
1666         ldlm_ibits_policy_local_to_wire(&lock->l_policy_data,
1667                                         &dlm_rep->lock_desc.l_policy_data);
1668         rc = ELDLM_OK;
1669         EXIT;
1670 out_put:
1671         LDLM_DEBUG(lock, "server-side convert handler END, rc = %d", rc);
1672         LDLM_LOCK_PUT(lock);
1673         req->rq_status = rc;
1674         return 0;
1675 }
1676
1677 /**
1678  * Cancel all the locks whose handles are packed into ldlm_request
1679  *
1680  * Called by server code expecting such combined cancel activity
1681  * requests.
1682  */
1683 int ldlm_request_cancel(struct ptlrpc_request *req,
1684                         const struct ldlm_request *dlm_req,
1685                         int first, enum lustre_at_flags flags)
1686 {
1687         struct ldlm_resource *res, *pres = NULL;
1688         struct ldlm_lock *lock;
1689         int i, count, done = 0;
1690         unsigned int size;
1691
1692         ENTRY;
1693
1694         size = req_capsule_get_size(&req->rq_pill, &RMF_DLM_REQ, RCL_CLIENT);
1695         if (size <= offsetof(struct ldlm_request, lock_handle) ||
1696             (size - offsetof(struct ldlm_request, lock_handle)) /
1697              sizeof(struct lustre_handle) < dlm_req->lock_count)
1698                 RETURN(0);
1699
1700         count = dlm_req->lock_count ? dlm_req->lock_count : 1;
1701         if (first >= count)
1702                 RETURN(0);
1703
1704         if (count == 1 && dlm_req->lock_handle[0].cookie == 0)
1705                 RETURN(0);
1706
1707         /*
1708          * There is no lock on the server at the replay time,
1709          * skip lock cancelling to make replay tests to pass.
1710          */
1711         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
1712                 RETURN(0);
1713
1714         LDLM_DEBUG_NOLOCK("server-side cancel handler START: %d locks, starting at %d",
1715                           count, first);
1716
1717         for (i = first; i < count; i++) {
1718                 lock = ldlm_handle2lock(&dlm_req->lock_handle[i]);
1719                 if (!lock) {
1720                         /* below message checked in replay-single.sh test_36 */
1721                         LDLM_DEBUG_NOLOCK("server-side cancel handler stale lock (cookie %llu)",
1722                                           dlm_req->lock_handle[i].cookie);
1723                         continue;
1724                 }
1725
1726                 res = lock->l_resource;
1727                 done++;
1728
1729                 /*
1730                  * This code is an optimization to only attempt lock
1731                  * granting on the resource (that could be CPU-expensive)
1732                  * after we are done cancelling lock in that resource.
1733                  */
1734                 if (res != pres) {
1735                         if (pres != NULL) {
1736                                 ldlm_reprocess_all(pres, NULL);
1737                                 LDLM_RESOURCE_DELREF(pres);
1738                                 ldlm_resource_putref(pres);
1739                         }
1740                         if (res != NULL) {
1741                                 ldlm_resource_getref(res);
1742                                 LDLM_RESOURCE_ADDREF(res);
1743
1744                                 if (!ldlm_is_discard_data(lock))
1745                                         ldlm_lvbo_update(res, lock,
1746                                                          NULL, 1);
1747                         }
1748                         pres = res;
1749                 }
1750
1751                 if ((flags & LATF_STATS) && ldlm_is_ast_sent(lock) &&
1752                     lock->l_blast_sent != 0) {
1753                         timeout_t delay = 0;
1754
1755                         if (ktime_get_real_seconds() > lock->l_blast_sent)
1756                                 delay = ktime_get_real_seconds() -
1757                                         lock->l_blast_sent;
1758                         LDLM_DEBUG(lock,
1759                                    "server cancels blocked lock after %ds",
1760                                    delay);
1761                         at_measured(&lock->l_export->exp_bl_lock_at, delay);
1762                 }
1763                 ldlm_lock_cancel(lock);
1764                 LDLM_LOCK_PUT(lock);
1765         }
1766         if (pres != NULL) {
1767                 ldlm_reprocess_all(pres, NULL);
1768                 LDLM_RESOURCE_DELREF(pres);
1769                 ldlm_resource_putref(pres);
1770         }
1771         LDLM_DEBUG_NOLOCK("server-side cancel handler END");
1772         RETURN(done);
1773 }
1774 EXPORT_SYMBOL(ldlm_request_cancel);
1775
1776 /**
1777  * Main LDLM entry point for server code to cancel locks.
1778  *
1779  * Typically gets called from service handler on LDLM_CANCEL opc.
1780  */
1781 int ldlm_handle_cancel(struct ptlrpc_request *req)
1782 {
1783         struct ldlm_request *dlm_req;
1784         int rc;
1785
1786         ENTRY;
1787
1788         dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1789         if (dlm_req == NULL) {
1790                 CDEBUG(D_INFO, "bad request buffer for cancel\n");
1791                 RETURN(-EFAULT);
1792         }
1793
1794         if (req_capsule_get_size(&req->rq_pill, &RMF_DLM_REQ, RCL_CLIENT) <
1795             offsetof(struct ldlm_request, lock_handle[1]))
1796                 RETURN(-EPROTO);
1797
1798         if (req->rq_export && req->rq_export->exp_nid_stats &&
1799             req->rq_export->exp_nid_stats->nid_ldlm_stats)
1800                 lprocfs_counter_incr(req->rq_export->exp_nid_stats->nid_ldlm_stats,
1801                                      LDLM_CANCEL - LDLM_FIRST_OPC);
1802
1803         rc = req_capsule_server_pack(&req->rq_pill);
1804         if (rc)
1805                 RETURN(rc);
1806
1807         if (!ldlm_request_cancel(req, dlm_req, 0, LATF_STATS))
1808                 req->rq_status = LUSTRE_ESTALE;
1809
1810         RETURN(ptlrpc_reply(req));
1811 }
1812 #endif /* HAVE_SERVER_SUPPORT */
1813
1814 /**
1815  * Server may pass additional information about blocking lock.
1816  * For IBITS locks it is conflicting bits which can be used for
1817  * lock convert instead of cancel.
1818  */
1819 void ldlm_bl_desc2lock(const struct ldlm_lock_desc *ld, struct ldlm_lock *lock)
1820 {
1821         struct ldlm_namespace *ns = ldlm_lock_to_ns(lock);
1822
1823         check_res_locked(lock->l_resource);
1824         if (ns_is_client(ns) && ld &&
1825             (lock->l_resource->lr_type == LDLM_IBITS)) {
1826                 /*
1827                  * Lock description contains policy of blocking lock,
1828                  * and its cancel_bits is used to pass conflicting bits.
1829                  * NOTE: ld can be NULL or can be not NULL but zeroed if
1830                  * passed from ldlm_bl_thread_blwi(), check below used bits
1831                  * in ld to make sure it is valid description.
1832                  *
1833                  * If server may replace lock resource keeping the same cookie,
1834                  * never use cancel bits from different resource, full cancel
1835                  * is to be used.
1836                  */
1837                 if (ld->l_policy_data.l_inodebits.cancel_bits &&
1838                     ldlm_res_eq(&ld->l_resource.lr_name,
1839                                 &lock->l_resource->lr_name) &&
1840                     !(ldlm_is_cbpending(lock) &&
1841                       lock->l_policy_data.l_inodebits.cancel_bits == 0)) {
1842                         /* always combine conflicting ibits */
1843                         lock->l_policy_data.l_inodebits.cancel_bits |=
1844                                 ld->l_policy_data.l_inodebits.cancel_bits;
1845                 } else {
1846                         /* If cancel_bits are not obtained or
1847                          * if the lock is already CBPENDING and
1848                          * has no cancel_bits set
1849                          * - the full lock is to be cancelled
1850                          */
1851                         lock->l_policy_data.l_inodebits.cancel_bits = 0;
1852                 }
1853         }
1854 }
1855
1856 /**
1857  * Callback handler for receiving incoming blocking ASTs.
1858  *
1859  * This can only happen on client side.
1860  */
1861 void ldlm_handle_bl_callback(struct ldlm_namespace *ns,
1862                              struct ldlm_lock_desc *ld, struct ldlm_lock *lock)
1863 {
1864         int do_ast;
1865
1866         ENTRY;
1867
1868         LDLM_DEBUG(lock, "client blocking AST callback handler");
1869
1870         lock_res_and_lock(lock);
1871
1872         /* get extra information from desc if any */
1873         ldlm_bl_desc2lock(ld, lock);
1874         ldlm_set_cbpending(lock);
1875
1876         do_ast = (!lock->l_readers && !lock->l_writers);
1877         unlock_res_and_lock(lock);
1878
1879         if (do_ast) {
1880                 CDEBUG(D_DLMTRACE,
1881                        "Lock %p already unused, calling callback (%p)\n",
1882                        lock, lock->l_blocking_ast);
1883                 if (lock->l_blocking_ast != NULL)
1884                         lock->l_blocking_ast(lock, ld, lock->l_ast_data,
1885                                              LDLM_CB_BLOCKING);
1886         } else {
1887                 CDEBUG(D_DLMTRACE,
1888                        "Lock %p is referenced, will be cancelled later\n",
1889                        lock);
1890         }
1891
1892         LDLM_DEBUG(lock, "client blocking callback handler END");
1893         LDLM_LOCK_RELEASE(lock);
1894         EXIT;
1895 }
1896
1897 static int ldlm_callback_reply(struct ptlrpc_request *req, int rc)
1898 {
1899         if (req->rq_no_reply)
1900                 return 0;
1901
1902         req->rq_status = rc;
1903         if (!req->rq_packed_final) {
1904                 rc = lustre_pack_reply(req, 1, NULL, NULL);
1905                 if (rc)
1906                         return rc;
1907         }
1908         return ptlrpc_reply(req);
1909 }
1910
1911 /**
1912  * Callback handler for receiving incoming completion ASTs.
1913  *
1914  * This only can happen on client side.
1915  */
1916 static int ldlm_handle_cp_callback(struct ptlrpc_request *req,
1917                                     struct ldlm_namespace *ns,
1918                                     struct ldlm_request *dlm_req,
1919                                     struct ldlm_lock *lock)
1920 {
1921         LIST_HEAD(ast_list);
1922         int lvb_len;
1923         int rc = 0;
1924
1925         ENTRY;
1926
1927         LDLM_DEBUG(lock, "client completion callback handler START");
1928
1929         if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_BL_CB_RACE)) {
1930                 long to = cfs_time_seconds(1);
1931
1932                 ldlm_callback_reply(req, 0);
1933
1934                 while (to > 0) {
1935                         to = schedule_timeout_interruptible(to);
1936                         if (ldlm_is_granted(lock) ||
1937                             ldlm_is_destroyed(lock))
1938                                 break;
1939                 }
1940         }
1941
1942         lvb_len = req_capsule_get_size(&req->rq_pill, &RMF_DLM_LVB, RCL_CLIENT);
1943         if (lvb_len < 0) {
1944                 LDLM_ERROR(lock, "Fail to get lvb_len, rc = %d", lvb_len);
1945                 GOTO(out, rc = lvb_len);
1946         } else if (lvb_len > 0) {
1947                 if (lock->l_lvb_len > 0) {
1948                         /* for extent lock, lvb contains ost_lvb{}. */
1949                         LASSERT(lock->l_lvb_data != NULL);
1950
1951                         if (unlikely(lock->l_lvb_len < lvb_len)) {
1952                                 LDLM_ERROR(lock,
1953                                            "Replied LVB is larger than expectation, expected = %d, replied = %d",
1954                                            lock->l_lvb_len, lvb_len);
1955                                 GOTO(out, rc = -EINVAL);
1956                         }
1957                 }
1958         }
1959
1960         lock_res_and_lock(lock);
1961
1962         if (!ldlm_res_eq(&dlm_req->lock_desc.l_resource.lr_name,
1963                          &lock->l_resource->lr_name)) {
1964                 ldlm_resource_unlink_lock(lock);
1965                 unlock_res_and_lock(lock);
1966                 rc = ldlm_lock_change_resource(ns, lock,
1967                                 &dlm_req->lock_desc.l_resource.lr_name);
1968                 if (rc < 0) {
1969                         LDLM_ERROR(lock, "Failed to allocate resource");
1970                         GOTO(out, rc);
1971                 }
1972                 LDLM_DEBUG(lock, "completion AST, new resource");
1973                 lock_res_and_lock(lock);
1974         }
1975
1976         if (ldlm_is_failed(lock)) {
1977                 unlock_res_and_lock(lock);
1978                 LDLM_LOCK_RELEASE(lock);
1979                 RETURN(-EINVAL);
1980         }
1981
1982         if (ldlm_is_destroyed(lock) ||
1983             ldlm_is_granted(lock)) {
1984                 /* b=11300: the lock has already been granted */
1985                 unlock_res_and_lock(lock);
1986                 LDLM_DEBUG(lock, "Double grant race happened");
1987                 GOTO(out, rc = 0);
1988         }
1989
1990         /*
1991          * If we receive the completion AST before the actual enqueue returned,
1992          * then we might need to switch lock modes, resources, or extents.
1993          */
1994         if (dlm_req->lock_desc.l_granted_mode != lock->l_req_mode) {
1995                 lock->l_req_mode = dlm_req->lock_desc.l_granted_mode;
1996                 LDLM_DEBUG(lock, "completion AST, new lock mode");
1997         }
1998
1999         if (lock->l_resource->lr_type != LDLM_PLAIN) {
2000                 ldlm_convert_policy_to_local(req->rq_export,
2001                                           dlm_req->lock_desc.l_resource.lr_type,
2002                                           &dlm_req->lock_desc.l_policy_data,
2003                                           &lock->l_policy_data);
2004                 LDLM_DEBUG(lock, "completion AST, new policy data");
2005         }
2006
2007         ldlm_resource_unlink_lock(lock);
2008
2009         if (dlm_req->lock_flags & LDLM_FL_AST_SENT) {
2010                 /*
2011                  * BL_AST locks are not needed in LRU.
2012                  * Let ldlm_cancel_lru() be fast.
2013                  */
2014                 ldlm_lock_remove_from_lru(lock);
2015                 ldlm_bl_desc2lock(&dlm_req->lock_desc, lock);
2016                 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_BL_AST;
2017                 LDLM_DEBUG(lock, "completion AST includes blocking AST");
2018         }
2019
2020         if (lock->l_lvb_len > 0) {
2021                 rc = ldlm_fill_lvb(lock, &req->rq_pill, RCL_CLIENT,
2022                                    lock->l_lvb_data, lvb_len);
2023                 if (rc < 0) {
2024                         unlock_res_and_lock(lock);
2025                         GOTO(out, rc);
2026                 }
2027         }
2028
2029         ldlm_grant_lock(lock, &ast_list);
2030         unlock_res_and_lock(lock);
2031
2032         LDLM_DEBUG(lock, "callback handler finished, about to run_ast_work");
2033
2034         /*
2035          * Let Enqueue to call osc_lock_upcall() and initialize
2036          * l_ast_data
2037          */
2038         OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_CP_ENQ_RACE, 2);
2039
2040         ldlm_run_ast_work(ns, &ast_list, LDLM_WORK_CP_AST);
2041
2042         LDLM_DEBUG_NOLOCK("client completion callback handler END (lock %p)",
2043                           lock);
2044         GOTO(out, rc);
2045
2046 out:
2047         if (rc < 0) {
2048                 lock_res_and_lock(lock);
2049                 ldlm_set_failed(lock);
2050                 unlock_res_and_lock(lock);
2051                 wake_up(&lock->l_waitq);
2052         }
2053         LDLM_LOCK_RELEASE(lock);
2054
2055         return 0;
2056 }
2057
2058 /**
2059  * Callback handler for receiving incoming glimpse ASTs.
2060  *
2061  * This only can happen on client side.  After handling the glimpse AST
2062  * we also consider dropping the lock here if it is unused locally for a
2063  * long time.
2064  */
2065 static void ldlm_handle_gl_callback(struct ptlrpc_request *req,
2066                                     struct ldlm_namespace *ns,
2067                                     struct ldlm_request *dlm_req,
2068                                     struct ldlm_lock *lock)
2069 {
2070         struct ldlm_lock_desc *ld = &dlm_req->lock_desc;
2071         int rc = -ENOSYS;
2072
2073         ENTRY;
2074
2075         LDLM_DEBUG(lock, "client glimpse AST callback handler");
2076
2077         if (lock->l_glimpse_ast != NULL)
2078                 rc = lock->l_glimpse_ast(lock, req);
2079
2080         if (req->rq_repmsg != NULL) {
2081                 ptlrpc_reply(req);
2082         } else {
2083                 req->rq_status = rc;
2084                 ptlrpc_error(req);
2085         }
2086
2087         lock_res_and_lock(lock);
2088         if (lock->l_granted_mode == LCK_PW &&
2089             !lock->l_readers && !lock->l_writers &&
2090             ktime_after(ktime_get(),
2091                         ktime_add(lock->l_last_used, ns->ns_dirty_age_limit))) {
2092                 unlock_res_and_lock(lock);
2093
2094                 /* For MDS glimpse it is always DOM lock, set corresponding
2095                  * cancel_bits to perform lock convert if needed
2096                  */
2097                 if (lock->l_resource->lr_type == LDLM_IBITS)
2098                         ld->l_policy_data.l_inodebits.cancel_bits =
2099                                                         MDS_INODELOCK_DOM;
2100                 if (ldlm_bl_to_thread_lock(ns, ld, lock))
2101                         ldlm_handle_bl_callback(ns, ld, lock);
2102
2103                 EXIT;
2104                 return;
2105         }
2106         unlock_res_and_lock(lock);
2107         LDLM_LOCK_RELEASE(lock);
2108         EXIT;
2109 }
2110
2111 static int __ldlm_bl_to_thread(struct ldlm_bl_work_item *blwi,
2112                                enum ldlm_cancel_flags cancel_flags)
2113 {
2114         struct ldlm_bl_pool *blp = ldlm_state->ldlm_bl_pool;
2115
2116         ENTRY;
2117
2118         spin_lock(&blp->blp_lock);
2119         if (blwi->blwi_lock &&
2120             ldlm_is_discard_data(blwi->blwi_lock)) {
2121                 /* add LDLM_FL_DISCARD_DATA requests to the priority list */
2122                 list_add_tail(&blwi->blwi_entry, &blp->blp_prio_list);
2123         } else {
2124                 /* other blocking callbacks are added to the regular list */
2125                 list_add_tail(&blwi->blwi_entry, &blp->blp_list);
2126         }
2127         spin_unlock(&blp->blp_lock);
2128
2129         wake_up(&blp->blp_waitq);
2130
2131         /*
2132          * can not check blwi->blwi_flags as blwi could be already freed in
2133          * LCF_ASYNC mode
2134          */
2135         if (!(cancel_flags & LCF_ASYNC))
2136                 wait_for_completion(&blwi->blwi_comp);
2137
2138         RETURN(0);
2139 }
2140
2141 static inline void init_blwi(struct ldlm_bl_work_item *blwi,
2142                              struct ldlm_namespace *ns,
2143                              struct ldlm_lock_desc *ld,
2144                              struct list_head *cancels, int count,
2145                              struct ldlm_lock *lock,
2146                              enum ldlm_cancel_flags cancel_flags)
2147 {
2148         init_completion(&blwi->blwi_comp);
2149         INIT_LIST_HEAD(&blwi->blwi_head);
2150
2151         if (current->flags & PF_MEMALLOC)
2152                 blwi->blwi_mem_pressure = 1;
2153
2154         blwi->blwi_ns = ns;
2155         blwi->blwi_flags = cancel_flags;
2156         if (ld != NULL)
2157                 blwi->blwi_ld = *ld;
2158         if (count) {
2159                 list_splice_init(cancels, &blwi->blwi_head);
2160                 blwi->blwi_count = count;
2161         } else {
2162                 blwi->blwi_lock = lock;
2163         }
2164 }
2165
2166 /**
2167  * Queues a list of locks \a cancels containing \a count locks
2168  * for later processing by a blocking thread.  If \a count is zero,
2169  * then the lock referenced as \a lock is queued instead.
2170  *
2171  * The blocking thread would then call ->l_blocking_ast callback in the lock.
2172  * If list addition fails an error is returned and caller is supposed to
2173  * call ->l_blocking_ast itself.
2174  */
2175 static int ldlm_bl_to_thread(struct ldlm_namespace *ns,
2176                              struct ldlm_lock_desc *ld,
2177                              struct ldlm_lock *lock,
2178                              struct list_head *cancels, int count,
2179                              enum ldlm_cancel_flags cancel_flags)
2180 {
2181         ENTRY;
2182
2183         if (cancels && count == 0)
2184                 RETURN(0);
2185
2186         if (cancel_flags & LCF_ASYNC) {
2187                 struct ldlm_bl_work_item *blwi;
2188
2189                 OBD_ALLOC(blwi, sizeof(*blwi));
2190                 if (blwi == NULL)
2191                         RETURN(-ENOMEM);
2192                 init_blwi(blwi, ns, ld, cancels, count, lock, cancel_flags);
2193
2194                 RETURN(__ldlm_bl_to_thread(blwi, cancel_flags));
2195         } else {
2196                 /*
2197                  * if it is synchronous call do minimum mem alloc, as it could
2198                  * be triggered from kernel shrinker
2199                  */
2200                 struct ldlm_bl_work_item blwi;
2201
2202                 memset(&blwi, 0, sizeof(blwi));
2203                 init_blwi(&blwi, ns, ld, cancels, count, lock, cancel_flags);
2204                 RETURN(__ldlm_bl_to_thread(&blwi, cancel_flags));
2205         }
2206 }
2207
2208
2209 int ldlm_bl_to_thread_lock(struct ldlm_namespace *ns, struct ldlm_lock_desc *ld,
2210                            struct ldlm_lock *lock)
2211 {
2212         return ldlm_bl_to_thread(ns, ld, lock, NULL, 0, LCF_ASYNC);
2213 }
2214
2215 int ldlm_bl_to_thread_list(struct ldlm_namespace *ns, struct ldlm_lock_desc *ld,
2216                            struct list_head *cancels, int count,
2217                            enum ldlm_cancel_flags cancel_flags)
2218 {
2219         return ldlm_bl_to_thread(ns, ld, NULL, cancels, count, cancel_flags);
2220 }
2221
2222 int ldlm_bl_to_thread_ns(struct ldlm_namespace *ns)
2223 {
2224         return ldlm_bl_to_thread(ns, NULL, NULL, NULL, 0, LCF_ASYNC);
2225 }
2226
2227 int ldlm_bl_thread_wakeup(void)
2228 {
2229         wake_up(&ldlm_state->ldlm_bl_pool->blp_waitq);
2230         return 0;
2231 }
2232
2233 /* Setinfo coming from Server (eg MDT) to Client (eg MDC)! */
2234 static int ldlm_handle_setinfo(struct ptlrpc_request *req)
2235 {
2236         struct obd_device *obd = req->rq_export->exp_obd;
2237         char *key;
2238         void *val;
2239         int keylen, vallen;
2240         int rc = -ENOSYS;
2241
2242         ENTRY;
2243
2244         DEBUG_REQ(D_HSM, req, "%s: handle setinfo", obd->obd_name);
2245
2246         req_capsule_set(&req->rq_pill, &RQF_OBD_SET_INFO);
2247
2248         key = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_KEY);
2249         if (key == NULL) {
2250                 DEBUG_REQ(D_IOCTL, req, "no set_info key");
2251                 RETURN(-EFAULT);
2252         }
2253         keylen = req_capsule_get_size(&req->rq_pill, &RMF_SETINFO_KEY,
2254                                       RCL_CLIENT);
2255         val = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_VAL);
2256         if (val == NULL) {
2257                 DEBUG_REQ(D_IOCTL, req, "no set_info val");
2258                 RETURN(-EFAULT);
2259         }
2260         vallen = req_capsule_get_size(&req->rq_pill, &RMF_SETINFO_VAL,
2261                                       RCL_CLIENT);
2262
2263         /* We are responsible for swabbing contents of val */
2264
2265         if (KEY_IS(KEY_HSM_COPYTOOL_SEND))
2266                 /* Pass it on to mdc (the "export" in this case) */
2267                 rc = obd_set_info_async(req->rq_svc_thread->t_env,
2268                                         req->rq_export,
2269                                         sizeof(KEY_HSM_COPYTOOL_SEND),
2270                                         KEY_HSM_COPYTOOL_SEND,
2271                                         vallen, val, NULL);
2272         else
2273                 DEBUG_REQ(D_WARNING, req, "ignoring unknown key '%s'", key);
2274
2275         return rc;
2276 }
2277
2278 static inline void ldlm_callback_errmsg(struct ptlrpc_request *req,
2279                                         const char *msg, int rc,
2280                                         const struct lustre_handle *handle)
2281 {
2282         DEBUG_REQ((req->rq_no_reply || rc) ? D_WARNING : D_DLMTRACE, req,
2283                   "%s, NID=%s lock=%#llx: rc = %d",
2284                   msg, libcfs_id2str(req->rq_peer),
2285                   handle ? handle->cookie : 0, rc);
2286         if (req->rq_no_reply)
2287                 CWARN("No reply was sent, maybe cause b=21636.\n");
2288         else if (rc)
2289                 CWARN("Send reply failed, maybe cause b=21636.\n");
2290 }
2291
2292 /* TODO: handle requests in a similar way as MDT: see mdt_handle_common() */
2293 static int ldlm_callback_handler(struct ptlrpc_request *req)
2294 {
2295         struct ldlm_namespace *ns;
2296         struct ldlm_request *dlm_req;
2297         struct ldlm_lock *lock;
2298         int rc;
2299
2300         ENTRY;
2301
2302         /*
2303          * Requests arrive in sender's byte order.  The ptlrpc service
2304          * handler has already checked and, if necessary, byte-swapped the
2305          * incoming request message body, but I am responsible for the
2306          * message buffers.
2307          */
2308
2309         /* do nothing for sec context finalize */
2310         if (lustre_msg_get_opc(req->rq_reqmsg) == SEC_CTX_FINI)
2311                 RETURN(0);
2312
2313         req_capsule_init(&req->rq_pill, req, RCL_SERVER);
2314
2315         if (req->rq_export == NULL) {
2316                 rc = ldlm_callback_reply(req, -ENOTCONN);
2317                 ldlm_callback_errmsg(req, "Operate on unconnected server",
2318                                      rc, NULL);
2319                 RETURN(0);
2320         }
2321
2322         LASSERT(req->rq_export != NULL);
2323         LASSERT(req->rq_export->exp_obd != NULL);
2324
2325         switch (lustre_msg_get_opc(req->rq_reqmsg)) {
2326         case LDLM_BL_CALLBACK:
2327                 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_BL_CALLBACK_NET)) {
2328                         if (cfs_fail_err)
2329                                 ldlm_callback_reply(req, -(int)cfs_fail_err);
2330                         RETURN(0);
2331                 }
2332                 break;
2333         case LDLM_CP_CALLBACK:
2334                 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CP_CALLBACK_NET))
2335                         RETURN(0);
2336                 break;
2337         case LDLM_GL_CALLBACK:
2338                 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_GL_CALLBACK_NET))
2339                         RETURN(0);
2340                 break;
2341         case LDLM_SET_INFO:
2342                 rc = ldlm_handle_setinfo(req);
2343                 ldlm_callback_reply(req, rc);
2344                 RETURN(0);
2345         default:
2346                 CERROR("unknown opcode %u\n",
2347                        lustre_msg_get_opc(req->rq_reqmsg));
2348                 ldlm_callback_reply(req, -EPROTO);
2349                 RETURN(0);
2350         }
2351
2352         ns = req->rq_export->exp_obd->obd_namespace;
2353         LASSERT(ns != NULL);
2354
2355         req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK);
2356
2357         dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2358         if (dlm_req == NULL) {
2359                 rc = ldlm_callback_reply(req, -EPROTO);
2360                 ldlm_callback_errmsg(req, "Operate without parameter", rc,
2361                                      NULL);
2362                 RETURN(0);
2363         }
2364
2365         /*
2366          * Force a known safe race, send a cancel to the server for a lock
2367          * which the server has already started a blocking callback on.
2368          */
2369         if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_BL_CB_RACE) &&
2370             lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK) {
2371                 rc = ldlm_cli_cancel(&dlm_req->lock_handle[0], 0);
2372                 if (rc < 0)
2373                         CERROR("ldlm_cli_cancel: %d\n", rc);
2374         }
2375
2376         lock = ldlm_handle2lock_long(&dlm_req->lock_handle[0], 0);
2377         if (!lock) {
2378                 CDEBUG(D_DLMTRACE,
2379                        "callback on lock %#llx - lock disappeared\n",
2380                        dlm_req->lock_handle[0].cookie);
2381                 rc = ldlm_callback_reply(req, -EINVAL);
2382                 ldlm_callback_errmsg(req, "Operate with invalid parameter", rc,
2383                                      &dlm_req->lock_handle[0]);
2384                 RETURN(0);
2385         }
2386
2387         if (ldlm_is_fail_loc(lock) &&
2388             lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK)
2389                 OBD_RACE(OBD_FAIL_LDLM_CP_BL_RACE);
2390
2391         /* Copy hints/flags (e.g. LDLM_FL_DISCARD_DATA) from AST. */
2392         lock_res_and_lock(lock);
2393         lock->l_flags |= ldlm_flags_from_wire(dlm_req->lock_flags &
2394                                               LDLM_FL_AST_MASK);
2395         if (lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK) {
2396                 /*
2397                  * If somebody cancels lock and cache is already dropped,
2398                  * or lock is failed before cp_ast received on client,
2399                  * we can tell the server we have no lock. Otherwise, we
2400                  * should send cancel after dropping the cache.
2401                  */
2402                 if ((ldlm_is_canceling(lock) && ldlm_is_bl_done(lock)) ||
2403                      ldlm_is_failed(lock)) {
2404                         LDLM_DEBUG(lock,
2405                                    "callback on lock %llx - lock disappeared",
2406                                    dlm_req->lock_handle[0].cookie);
2407                         unlock_res_and_lock(lock);
2408                         LDLM_LOCK_RELEASE(lock);
2409                         rc = ldlm_callback_reply(req, -EINVAL);
2410                         ldlm_callback_errmsg(req, "Operate on stale lock", rc,
2411                                              &dlm_req->lock_handle[0]);
2412                         RETURN(0);
2413                 }
2414                 /*
2415                  * BL_AST locks are not needed in LRU.
2416                  * Let ldlm_cancel_lru() be fast.
2417                  */
2418                 ldlm_lock_remove_from_lru(lock);
2419                 ldlm_set_bl_ast(lock);
2420         }
2421         if (lock->l_remote_handle.cookie == 0)
2422                 lock->l_remote_handle = dlm_req->lock_handle[1];
2423         unlock_res_and_lock(lock);
2424
2425         /*
2426          * We want the ost thread to get this reply so that it can respond
2427          * to ost requests (write cache writeback) that might be triggered
2428          * in the callback.
2429          *
2430          * But we'd also like to be able to indicate in the reply that we're
2431          * cancelling right now, because it's unused, or have an intent result
2432          * in the reply, so we might have to push the responsibility for sending
2433          * the reply down into the AST handlers, alas.
2434          */
2435
2436         switch (lustre_msg_get_opc(req->rq_reqmsg)) {
2437         case LDLM_BL_CALLBACK:
2438                 CDEBUG(D_INODE, "blocking ast\n");
2439                 req_capsule_extend(&req->rq_pill, &RQF_LDLM_BL_CALLBACK);
2440                 if (!ldlm_is_cancel_on_block(lock)) {
2441                         rc = ldlm_callback_reply(req, 0);
2442                         if (req->rq_no_reply || rc)
2443                                 ldlm_callback_errmsg(req, "Normal process", rc,
2444                                                      &dlm_req->lock_handle[0]);
2445                 }
2446                 if (ldlm_bl_to_thread_lock(ns, &dlm_req->lock_desc, lock))
2447                         ldlm_handle_bl_callback(ns, &dlm_req->lock_desc, lock);
2448                 break;
2449         case LDLM_CP_CALLBACK:
2450                 CDEBUG(D_INODE, "completion ast\n");
2451                 req_capsule_extend(&req->rq_pill, &RQF_LDLM_CP_CALLBACK);
2452                 rc = ldlm_handle_cp_callback(req, ns, dlm_req, lock);
2453                 if (!OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_BL_CB_RACE))
2454                         ldlm_callback_reply(req, rc);
2455                 break;
2456         case LDLM_GL_CALLBACK:
2457                 CDEBUG(D_INODE, "glimpse ast\n");
2458                 req_capsule_extend(&req->rq_pill, &RQF_LDLM_GL_CALLBACK);
2459                 ldlm_handle_gl_callback(req, ns, dlm_req, lock);
2460                 break;
2461         default:
2462                 LBUG(); /* checked above */
2463         }
2464
2465         RETURN(0);
2466 }
2467
2468 #ifdef HAVE_SERVER_SUPPORT
2469 /**
2470  * Main handler for canceld thread.
2471  *
2472  * Separated into its own thread to avoid deadlocks.
2473  */
2474 static int ldlm_cancel_handler(struct ptlrpc_request *req)
2475 {
2476         int rc;
2477
2478         ENTRY;
2479
2480         /*
2481          * Requests arrive in sender's byte order.  The ptlrpc service
2482          * handler has already checked and, if necessary, byte-swapped the
2483          * incoming request message body, but I am responsible for the
2484          * message buffers.
2485          */
2486
2487         req_capsule_init(&req->rq_pill, req, RCL_SERVER);
2488
2489         if (req->rq_export == NULL) {
2490                 struct ldlm_request *dlm_req;
2491
2492                 CERROR("%s from %s arrived at %llu with bad export cookie %llu\n",
2493                        ll_opcode2str(lustre_msg_get_opc(req->rq_reqmsg)),
2494                        libcfs_nid2str(req->rq_peer.nid),
2495                        (unsigned long long)req->rq_arrival_time.tv_sec,
2496                        lustre_msg_get_handle(req->rq_reqmsg)->cookie);
2497
2498                 if (lustre_msg_get_opc(req->rq_reqmsg) == LDLM_CANCEL) {
2499                         req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK);
2500                         dlm_req = req_capsule_client_get(&req->rq_pill,
2501                                                          &RMF_DLM_REQ);
2502                         if (dlm_req != NULL)
2503                                 ldlm_lock_dump_handle(D_ERROR,
2504                                                       &dlm_req->lock_handle[0]);
2505                 }
2506                 ldlm_callback_reply(req, -ENOTCONN);
2507                 RETURN(0);
2508         }
2509
2510         switch (lustre_msg_get_opc(req->rq_reqmsg)) {
2511         /* XXX FIXME move this back to mds/handler.c, b=249 */
2512         case LDLM_CANCEL:
2513                 req_capsule_set(&req->rq_pill, &RQF_LDLM_CANCEL);
2514                 CDEBUG(D_INODE, "cancel\n");
2515                 if (CFS_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_NET) ||
2516                     CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_CANCEL_RESEND) ||
2517                     CFS_FAIL_CHECK(OBD_FAIL_LDLM_BL_EVICT))
2518                         RETURN(0);
2519                 rc = ldlm_handle_cancel(req);
2520                 break;
2521         case LDLM_CONVERT:
2522         {
2523                 struct ldlm_request *dlm_req;
2524
2525                 req_capsule_set(&req->rq_pill, &RQF_LDLM_CONVERT);
2526                 CDEBUG(D_INODE, "convert\n");
2527
2528                 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2529                 if (dlm_req == NULL) {
2530                         CDEBUG(D_INFO, "bad request buffer for cancel\n");
2531                         rc = ldlm_callback_reply(req, -EPROTO);
2532                 } else {
2533                         req->rq_status = ldlm_handle_convert0(req, dlm_req);
2534                         rc = ptlrpc_reply(req);
2535                 }
2536                 break;
2537         }
2538         default:
2539                 CERROR("invalid opcode %d\n",
2540                        lustre_msg_get_opc(req->rq_reqmsg));
2541                 req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK);
2542                 rc = ldlm_callback_reply(req, -EINVAL);
2543         }
2544
2545         RETURN(rc);
2546 }
2547
2548 static int ldlm_cancel_hpreq_lock_match(struct ptlrpc_request *req,
2549                                         struct ldlm_lock *lock)
2550 {
2551         struct ldlm_request *dlm_req;
2552         struct lustre_handle lockh;
2553         int rc = 0;
2554         int i;
2555
2556         ENTRY;
2557
2558         dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2559         if (dlm_req == NULL)
2560                 RETURN(0);
2561
2562         ldlm_lock2handle(lock, &lockh);
2563         for (i = 0; i < dlm_req->lock_count; i++) {
2564                 if (lustre_handle_equal(&dlm_req->lock_handle[i],
2565                                         &lockh)) {
2566                         DEBUG_REQ(D_RPCTRACE, req,
2567                                   "Prio raised by lock %#llx", lockh.cookie);
2568                         rc = 1;
2569                         break;
2570                 }
2571         }
2572
2573         RETURN(rc);
2574 }
2575
2576 static int ldlm_cancel_hpreq_check(struct ptlrpc_request *req)
2577 {
2578         struct ldlm_request *dlm_req;
2579         int rc = 0;
2580         int i;
2581         unsigned int size;
2582
2583         ENTRY;
2584
2585         /* no prolong in recovery */
2586         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
2587                 RETURN(0);
2588
2589         dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2590         if (dlm_req == NULL)
2591                 RETURN(-EFAULT);
2592
2593         size = req_capsule_get_size(&req->rq_pill, &RMF_DLM_REQ, RCL_CLIENT);
2594         if (size <= offsetof(struct ldlm_request, lock_handle) ||
2595             (size - offsetof(struct ldlm_request, lock_handle)) /
2596              sizeof(struct lustre_handle) < dlm_req->lock_count)
2597                 RETURN(-EPROTO);
2598
2599         for (i = 0; i < dlm_req->lock_count; i++) {
2600                 struct ldlm_lock *lock;
2601
2602                 lock = ldlm_handle2lock(&dlm_req->lock_handle[i]);
2603                 if (lock == NULL)
2604                         continue;
2605
2606                 rc = ldlm_is_ast_sent(lock) ? 1 : 0;
2607                 if (rc)
2608                         LDLM_DEBUG(lock, "hpreq cancel/convert lock");
2609                 LDLM_LOCK_PUT(lock);
2610
2611                 if (rc)
2612                         break;
2613         }
2614
2615         RETURN(rc);
2616 }
2617
2618 static struct ptlrpc_hpreq_ops ldlm_cancel_hpreq_ops = {
2619         .hpreq_lock_match = ldlm_cancel_hpreq_lock_match,
2620         .hpreq_check      = ldlm_cancel_hpreq_check,
2621         .hpreq_fini       = NULL,
2622 };
2623
2624 static int ldlm_hpreq_handler(struct ptlrpc_request *req)
2625 {
2626         ENTRY;
2627
2628         req_capsule_init(&req->rq_pill, req, RCL_SERVER);
2629
2630         if (req->rq_export == NULL)
2631                 RETURN(0);
2632
2633         if (LDLM_CANCEL == lustre_msg_get_opc(req->rq_reqmsg)) {
2634                 req_capsule_set(&req->rq_pill, &RQF_LDLM_CANCEL);
2635                 req->rq_ops = &ldlm_cancel_hpreq_ops;
2636         } else if (LDLM_CONVERT == lustre_msg_get_opc(req->rq_reqmsg)) {
2637                 req_capsule_set(&req->rq_pill, &RQF_LDLM_CONVERT);
2638                 req->rq_ops = &ldlm_cancel_hpreq_ops;
2639         }
2640         RETURN(0);
2641 }
2642
2643 static int ldlm_revoke_lock_cb(struct cfs_hash *hs, struct cfs_hash_bd *bd,
2644                                struct hlist_node *hnode, void *data)
2645
2646 {
2647         struct list_head *rpc_list = data;
2648         struct ldlm_lock *lock = cfs_hash_object(hs, hnode);
2649
2650         lock_res_and_lock(lock);
2651
2652         if (!ldlm_is_granted(lock)) {
2653                 unlock_res_and_lock(lock);
2654                 return 0;
2655         }
2656
2657         LASSERT(lock->l_resource);
2658         if (lock->l_resource->lr_type != LDLM_IBITS &&
2659             lock->l_resource->lr_type != LDLM_PLAIN) {
2660                 unlock_res_and_lock(lock);
2661                 return 0;
2662         }
2663
2664         if (ldlm_is_ast_sent(lock)) {
2665                 unlock_res_and_lock(lock);
2666                 return 0;
2667         }
2668
2669         LASSERT(lock->l_blocking_ast);
2670         LASSERT(!lock->l_blocking_lock);
2671
2672         ldlm_set_ast_sent(lock);
2673         if (lock->l_export && lock->l_export->exp_lock_hash) {
2674                 /*
2675                  * NB: it's safe to call cfs_hash_del() even lock isn't
2676                  * in exp_lock_hash.
2677                  */
2678                 /*
2679                  * In the function below, .hs_keycmp resolves to
2680                  * ldlm_export_lock_keycmp()
2681                  */
2682                 /* coverity[overrun-buffer-val] */
2683                 cfs_hash_del(lock->l_export->exp_lock_hash,
2684                              &lock->l_remote_handle, &lock->l_exp_hash);
2685         }
2686
2687         list_add_tail(&lock->l_rk_ast, rpc_list);
2688         LDLM_LOCK_GET(lock);
2689
2690         unlock_res_and_lock(lock);
2691         return 0;
2692 }
2693
2694 void ldlm_revoke_export_locks(struct obd_export *exp)
2695 {
2696         LIST_HEAD(rpc_list);
2697
2698         ENTRY;
2699
2700         cfs_hash_for_each_nolock(exp->exp_lock_hash,
2701                                  ldlm_revoke_lock_cb, &rpc_list, 0);
2702         ldlm_run_ast_work(exp->exp_obd->obd_namespace, &rpc_list,
2703                           LDLM_WORK_REVOKE_AST);
2704
2705         EXIT;
2706 }
2707 EXPORT_SYMBOL(ldlm_revoke_export_locks);
2708 #endif /* HAVE_SERVER_SUPPORT */
2709
2710 static int ldlm_bl_get_work(struct ldlm_bl_pool *blp,
2711                             struct ldlm_bl_work_item **p_blwi,
2712                             struct obd_export **p_exp)
2713 {
2714         struct ldlm_bl_work_item *blwi = NULL;
2715         static unsigned int num_bl;
2716         static unsigned int num_stale;
2717         int num_th = atomic_read(&blp->blp_num_threads);
2718
2719         *p_exp = obd_stale_export_get();
2720
2721         spin_lock(&blp->blp_lock);
2722         if (*p_exp != NULL) {
2723                 if (num_th == 1 || ++num_stale < num_th) {
2724                         spin_unlock(&blp->blp_lock);
2725                         return 1;
2726                 }
2727                 num_stale = 0;
2728         }
2729
2730         /* process a request from the blp_list at least every blp_num_threads */
2731         if (!list_empty(&blp->blp_list) &&
2732             (list_empty(&blp->blp_prio_list) || num_bl == 0))
2733                 blwi = list_entry(blp->blp_list.next,
2734                                   struct ldlm_bl_work_item, blwi_entry);
2735         else
2736                 if (!list_empty(&blp->blp_prio_list))
2737                         blwi = list_entry(blp->blp_prio_list.next,
2738                                           struct ldlm_bl_work_item,
2739                                           blwi_entry);
2740
2741         if (blwi) {
2742                 if (++num_bl >= num_th)
2743                         num_bl = 0;
2744                 list_del(&blwi->blwi_entry);
2745         }
2746         spin_unlock(&blp->blp_lock);
2747         *p_blwi = blwi;
2748
2749         if (*p_exp != NULL && *p_blwi != NULL) {
2750                 obd_stale_export_put(*p_exp);
2751                 *p_exp = NULL;
2752         }
2753
2754         return (*p_blwi != NULL || *p_exp != NULL) ? 1 : 0;
2755 }
2756
2757 /* This only contains temporary data until the thread starts */
2758 struct ldlm_bl_thread_data {
2759         struct ldlm_bl_pool     *bltd_blp;
2760         struct completion       bltd_comp;
2761         int                     bltd_num;
2762 };
2763
2764 static int ldlm_bl_thread_main(void *arg);
2765
2766 static int ldlm_bl_thread_start(struct ldlm_bl_pool *blp, bool check_busy)
2767 {
2768         struct ldlm_bl_thread_data bltd = { .bltd_blp = blp };
2769         struct task_struct *task;
2770
2771         init_completion(&bltd.bltd_comp);
2772
2773         bltd.bltd_num = atomic_inc_return(&blp->blp_num_threads);
2774         if (bltd.bltd_num >= blp->blp_max_threads) {
2775                 atomic_dec(&blp->blp_num_threads);
2776                 return 0;
2777         }
2778
2779         LASSERTF(bltd.bltd_num > 0, "thread num:%d\n", bltd.bltd_num);
2780         if (check_busy &&
2781             atomic_read(&blp->blp_busy_threads) < (bltd.bltd_num - 1)) {
2782                 atomic_dec(&blp->blp_num_threads);
2783                 return 0;
2784         }
2785
2786         task = kthread_run(ldlm_bl_thread_main, &bltd, "ldlm_bl_%02d",
2787                            bltd.bltd_num);
2788         if (IS_ERR(task)) {
2789                 CERROR("cannot start LDLM thread ldlm_bl_%02d: rc %ld\n",
2790                        bltd.bltd_num, PTR_ERR(task));
2791                 atomic_dec(&blp->blp_num_threads);
2792                 return PTR_ERR(task);
2793         }
2794         wait_for_completion(&bltd.bltd_comp);
2795
2796         return 0;
2797 }
2798
2799 /* Not fatal if racy and have a few too many threads */
2800 static int ldlm_bl_thread_need_create(struct ldlm_bl_pool *blp,
2801                                       struct ldlm_bl_work_item *blwi)
2802 {
2803         if (atomic_read(&blp->blp_num_threads) >= blp->blp_max_threads)
2804                 return 0;
2805
2806         if (atomic_read(&blp->blp_busy_threads) <
2807             atomic_read(&blp->blp_num_threads))
2808                 return 0;
2809
2810         if (blwi != NULL && (blwi->blwi_ns == NULL ||
2811                              blwi->blwi_mem_pressure))
2812                 return 0;
2813
2814         return 1;
2815 }
2816
2817 static int ldlm_bl_thread_blwi(struct ldlm_bl_pool *blp,
2818                                struct ldlm_bl_work_item *blwi)
2819 {
2820         /* '1' for consistency with code that checks !mpflag to restore */
2821         unsigned int mpflags = 1;
2822
2823         ENTRY;
2824
2825         if (blwi->blwi_ns == NULL)
2826                 /* added by ldlm_cleanup() */
2827                 RETURN(LDLM_ITER_STOP);
2828
2829         if (blwi->blwi_mem_pressure)
2830                 mpflags = memalloc_noreclaim_save();
2831
2832         OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_PAUSE_CANCEL2, 4);
2833
2834         if (blwi->blwi_count) {
2835                 int count;
2836                 /*
2837                  * The special case when we cancel locks in lru
2838                  * asynchronously, we pass the list of locks here.
2839                  * Thus locks are marked LDLM_FL_CANCELING, but NOT
2840                  * canceled locally yet.
2841                  */
2842                 count = ldlm_cli_cancel_list_local(&blwi->blwi_head,
2843                                                    blwi->blwi_count,
2844                                                    LCF_BL_AST);
2845                 ldlm_cli_cancel_list(&blwi->blwi_head, count, NULL,
2846                                      blwi->blwi_flags);
2847         } else if (blwi->blwi_lock) {
2848                 ldlm_handle_bl_callback(blwi->blwi_ns, &blwi->blwi_ld,
2849                                         blwi->blwi_lock);
2850         } else {
2851                 ldlm_pool_recalc(&blwi->blwi_ns->ns_pool, true);
2852                 spin_lock(&blwi->blwi_ns->ns_lock);
2853                 blwi->blwi_ns->ns_rpc_recalc = 0;
2854                 spin_unlock(&blwi->blwi_ns->ns_lock);
2855                 ldlm_namespace_put(blwi->blwi_ns);
2856         }
2857
2858         if (blwi->blwi_mem_pressure)
2859                 memalloc_noreclaim_restore(mpflags);
2860
2861         if (blwi->blwi_flags & LCF_ASYNC)
2862                 OBD_FREE(blwi, sizeof(*blwi));
2863         else
2864                 complete(&blwi->blwi_comp);
2865
2866         RETURN(0);
2867 }
2868
2869 /**
2870  * Cancel stale locks on export. Cancel blocked locks first.
2871  * If the given export has blocked locks, the next in the list may have
2872  * them too, thus cancel not blocked locks only if the current export has
2873  * no blocked locks.
2874  **/
2875 static int ldlm_bl_thread_exports(struct ldlm_bl_pool *blp,
2876                                   struct obd_export *exp)
2877 {
2878         int num;
2879
2880         ENTRY;
2881
2882         OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_BL_EVICT, 4);
2883
2884         num = ldlm_export_cancel_blocked_locks(exp);
2885         if (num == 0)
2886                 ldlm_export_cancel_locks(exp);
2887
2888         obd_stale_export_put(exp);
2889
2890         RETURN(0);
2891 }
2892
2893
2894 /**
2895  * Main blocking requests processing thread.
2896  *
2897  * Callers put locks into its queue by calling ldlm_bl_to_thread.
2898  * This thread in the end ends up doing actual call to ->l_blocking_ast
2899  * for queued locks.
2900  */
2901 static int ldlm_bl_thread_main(void *arg)
2902 {
2903         struct lu_env *env;
2904         struct ldlm_bl_pool *blp;
2905         struct ldlm_bl_thread_data *bltd = arg;
2906         int rc;
2907
2908         ENTRY;
2909
2910         OBD_ALLOC_PTR(env);
2911         if (!env)
2912                 RETURN(-ENOMEM);
2913         rc = lu_env_init(env, LCT_DT_THREAD);
2914         if (rc)
2915                 GOTO(out_env, rc);
2916         rc = lu_env_add(env);
2917         if (rc)
2918                 GOTO(out_env_fini, rc);
2919
2920         blp = bltd->bltd_blp;
2921
2922         complete(&bltd->bltd_comp);
2923         /* cannot use bltd after this, it is only on caller's stack */
2924
2925         while (1) {
2926                 struct ldlm_bl_work_item *blwi = NULL;
2927                 struct obd_export *exp = NULL;
2928                 int rc;
2929
2930                 rc = ldlm_bl_get_work(blp, &blwi, &exp);
2931
2932                 if (rc == 0)
2933                         wait_event_idle_exclusive(blp->blp_waitq,
2934                                                   ldlm_bl_get_work(blp, &blwi,
2935                                                                    &exp));
2936                 atomic_inc(&blp->blp_busy_threads);
2937
2938                 if (ldlm_bl_thread_need_create(blp, blwi))
2939                         /* discard the return value, we tried */
2940                         ldlm_bl_thread_start(blp, true);
2941
2942                 if (exp)
2943                         rc = ldlm_bl_thread_exports(blp, exp);
2944                 else if (blwi)
2945                         rc = ldlm_bl_thread_blwi(blp, blwi);
2946
2947                 atomic_dec(&blp->blp_busy_threads);
2948
2949                 if (rc == LDLM_ITER_STOP)
2950                         break;
2951
2952                 /*
2953                  * If there are many namespaces, we will not sleep waiting for
2954                  * work, and must do a cond_resched to avoid holding the CPU
2955                  * for too long
2956                  */
2957                 cond_resched();
2958         }
2959
2960         atomic_dec(&blp->blp_num_threads);
2961         complete(&blp->blp_comp);
2962
2963         lu_env_remove(env);
2964 out_env_fini:
2965         lu_env_fini(env);
2966 out_env:
2967         OBD_FREE_PTR(env);
2968         RETURN(rc);
2969 }
2970
2971
2972 static int ldlm_setup(void);
2973 static int ldlm_cleanup(void);
2974
2975 int ldlm_get_ref(void)
2976 {
2977         int rc = 0;
2978
2979         ENTRY;
2980         mutex_lock(&ldlm_ref_mutex);
2981         if (++ldlm_refcount == 1) {
2982                 rc = ldlm_setup();
2983                 if (rc)
2984                         ldlm_refcount--;
2985         }
2986         mutex_unlock(&ldlm_ref_mutex);
2987
2988         RETURN(rc);
2989 }
2990
2991 void ldlm_put_ref(void)
2992 {
2993         ENTRY;
2994         mutex_lock(&ldlm_ref_mutex);
2995         if (ldlm_refcount == 1) {
2996                 int rc = ldlm_cleanup();
2997
2998                 if (rc)
2999                         CERROR("ldlm_cleanup failed: %d\n", rc);
3000                 else
3001                         ldlm_refcount--;
3002         } else {
3003                 ldlm_refcount--;
3004         }
3005         mutex_unlock(&ldlm_ref_mutex);
3006
3007         EXIT;
3008 }
3009
3010 /*
3011  * Export handle<->lock hash operations.
3012  */
3013 static unsigned
3014 ldlm_export_lock_hash(struct cfs_hash *hs, const void *key, unsigned int mask)
3015 {
3016         return cfs_hash_u64_hash(((struct lustre_handle *)key)->cookie, mask);
3017 }
3018
3019 static void *
3020 ldlm_export_lock_key(struct hlist_node *hnode)
3021 {
3022         struct ldlm_lock *lock;
3023
3024         lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
3025         return &lock->l_remote_handle;
3026 }
3027
3028 static void
3029 ldlm_export_lock_keycpy(struct hlist_node *hnode, void *key)
3030 {
3031         struct ldlm_lock     *lock;
3032
3033         lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
3034         lock->l_remote_handle = *(struct lustre_handle *)key;
3035 }
3036
3037 static int
3038 ldlm_export_lock_keycmp(const void *key, struct hlist_node *hnode)
3039 {
3040         return lustre_handle_equal(ldlm_export_lock_key(hnode), key);
3041 }
3042
3043 static void *
3044 ldlm_export_lock_object(struct hlist_node *hnode)
3045 {
3046         return hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
3047 }
3048
3049 static void
3050 ldlm_export_lock_get(struct cfs_hash *hs, struct hlist_node *hnode)
3051 {
3052         struct ldlm_lock *lock;
3053
3054         lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
3055         LDLM_LOCK_GET(lock);
3056 }
3057
3058 static void
3059 ldlm_export_lock_put(struct cfs_hash *hs, struct hlist_node *hnode)
3060 {
3061         struct ldlm_lock *lock;
3062
3063         lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
3064         LDLM_LOCK_RELEASE(lock);
3065 }
3066
3067 static struct cfs_hash_ops ldlm_export_lock_ops = {
3068         .hs_hash        = ldlm_export_lock_hash,
3069         .hs_key         = ldlm_export_lock_key,
3070         .hs_keycmp      = ldlm_export_lock_keycmp,
3071         .hs_keycpy      = ldlm_export_lock_keycpy,
3072         .hs_object      = ldlm_export_lock_object,
3073         .hs_get         = ldlm_export_lock_get,
3074         .hs_put         = ldlm_export_lock_put,
3075         .hs_put_locked  = ldlm_export_lock_put,
3076 };
3077
3078 int ldlm_init_export(struct obd_export *exp)
3079 {
3080         int rc;
3081
3082         ENTRY;
3083
3084         exp->exp_lock_hash =
3085                 cfs_hash_create(obd_uuid2str(&exp->exp_client_uuid),
3086                                 HASH_EXP_LOCK_CUR_BITS,
3087                                 HASH_EXP_LOCK_MAX_BITS,
3088                                 HASH_EXP_LOCK_BKT_BITS, 0,
3089                                 CFS_HASH_MIN_THETA, CFS_HASH_MAX_THETA,
3090                                 &ldlm_export_lock_ops,
3091                                 CFS_HASH_DEFAULT | CFS_HASH_REHASH_KEY |
3092                                 CFS_HASH_NBLK_CHANGE);
3093
3094         if (!exp->exp_lock_hash)
3095                 RETURN(-ENOMEM);
3096
3097         rc = ldlm_init_flock_export(exp);
3098         if (rc)
3099                 GOTO(err, rc);
3100
3101         RETURN(0);
3102 err:
3103         ldlm_destroy_export(exp);
3104         RETURN(rc);
3105 }
3106 EXPORT_SYMBOL(ldlm_init_export);
3107
3108 void ldlm_destroy_export(struct obd_export *exp)
3109 {
3110         ENTRY;
3111         cfs_hash_putref(exp->exp_lock_hash);
3112         exp->exp_lock_hash = NULL;
3113
3114         ldlm_destroy_flock_export(exp);
3115         EXIT;
3116 }
3117 EXPORT_SYMBOL(ldlm_destroy_export);
3118
3119 static ssize_t cancel_unused_locks_before_replay_show(struct kobject *kobj,
3120                                                       struct attribute *attr,
3121                                                       char *buf)
3122 {
3123         return sprintf(buf, "%d\n", ldlm_cancel_unused_locks_before_replay);
3124 }
3125
3126 static ssize_t cancel_unused_locks_before_replay_store(struct kobject *kobj,
3127                                                        struct attribute *attr,
3128                                                        const char *buffer,
3129                                                        size_t count)
3130 {
3131         int rc;
3132         unsigned long val;
3133
3134         rc = kstrtoul(buffer, 10, &val);
3135         if (rc)
3136                 return rc;
3137
3138         ldlm_cancel_unused_locks_before_replay = val;
3139
3140         return count;
3141 }
3142 LUSTRE_RW_ATTR(cancel_unused_locks_before_replay);
3143
3144 static struct attribute *ldlm_attrs[] = {
3145         &lustre_attr_cancel_unused_locks_before_replay.attr,
3146         NULL,
3147 };
3148
3149 static struct attribute_group ldlm_attr_group = {
3150         .attrs = ldlm_attrs,
3151 };
3152
3153 static int ldlm_setup(void)
3154 {
3155         static struct ptlrpc_service_conf       conf;
3156         struct ldlm_bl_pool                    *blp = NULL;
3157 #ifdef HAVE_SERVER_SUPPORT
3158         struct task_struct *task;
3159 #endif /* HAVE_SERVER_SUPPORT */
3160         int i;
3161         int rc = 0;
3162
3163         ENTRY;
3164
3165         if (ldlm_state != NULL)
3166                 RETURN(-EALREADY);
3167
3168         OBD_ALLOC(ldlm_state, sizeof(*ldlm_state));
3169         if (ldlm_state == NULL)
3170                 RETURN(-ENOMEM);
3171
3172         ldlm_kobj = kobject_create_and_add("ldlm", &lustre_kset->kobj);
3173         if (!ldlm_kobj)
3174                 GOTO(out, -ENOMEM);
3175
3176         rc = sysfs_create_group(ldlm_kobj, &ldlm_attr_group);
3177         if (rc)
3178                 GOTO(out, rc);
3179
3180         ldlm_ns_kset = kset_create_and_add("namespaces", NULL, ldlm_kobj);
3181         if (!ldlm_ns_kset)
3182                 GOTO(out, -ENOMEM);
3183
3184         ldlm_svc_kset = kset_create_and_add("services", NULL, ldlm_kobj);
3185         if (!ldlm_svc_kset)
3186                 GOTO(out, -ENOMEM);
3187
3188         rc = ldlm_debugfs_setup();
3189         if (rc != 0)
3190                 GOTO(out, rc);
3191
3192         memset(&conf, 0, sizeof(conf));
3193         conf = (typeof(conf)) {
3194                 .psc_name               = "ldlm_cbd",
3195                 .psc_watchdog_factor    = 2,
3196                 .psc_buf                = {
3197                         .bc_nbufs               = LDLM_CLIENT_NBUFS,
3198                         .bc_buf_size            = LDLM_BUFSIZE,
3199                         .bc_req_max_size        = LDLM_MAXREQSIZE,
3200                         .bc_rep_max_size        = LDLM_MAXREPSIZE,
3201                         .bc_req_portal          = LDLM_CB_REQUEST_PORTAL,
3202                         .bc_rep_portal          = LDLM_CB_REPLY_PORTAL,
3203                 },
3204                 .psc_thr                = {
3205                         .tc_thr_name            = "ldlm_cb",
3206                         .tc_thr_factor          = LDLM_THR_FACTOR,
3207                         .tc_nthrs_init          = LDLM_NTHRS_INIT,
3208                         .tc_nthrs_base          = LDLM_NTHRS_BASE,
3209                         .tc_nthrs_max           = LDLM_NTHRS_MAX,
3210                         .tc_nthrs_user          = ldlm_num_threads,
3211                         .tc_cpu_bind            = ldlm_cpu_bind,
3212                         .tc_ctx_tags            = LCT_MD_THREAD | LCT_DT_THREAD,
3213                 },
3214                 .psc_cpt                = {
3215                         .cc_pattern             = ldlm_cpts,
3216                         .cc_affinity            = true,
3217                 },
3218                 .psc_ops                = {
3219                         .so_req_handler         = ldlm_callback_handler,
3220      &n