Whamcloud - gitweb
Branch HEAD
[fs/lustre-release.git] / lustre / ldlm / ldlm_lockd.c
1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2  * vim:expandtab:shiftwidth=8:tabstop=8:
3  *
4  * Copyright (C) 2002-2004 Cluster File Systems, Inc.
5  *   Author: Peter Braam <braam@clusterfs.com>
6  *   Author: Phil Schwan <phil@clusterfs.com>
7  *
8  *   This file is part of the Lustre file system, http://www.lustre.org
9  *   Lustre is a trademark of Cluster File Systems, Inc.
10  *
11  *   You may have signed or agreed to another license before downloading
12  *   this software.  If so, you are bound by the terms and conditions
13  *   of that agreement, and the following does not apply to you.  See the
14  *   LICENSE file included with this distribution for more information.
15  *
16  *   If you did not agree to a different license, then this copy of Lustre
17  *   is open source software; you can redistribute it and/or modify it
18  *   under the terms of version 2 of the GNU General Public License as
19  *   published by the Free Software Foundation.
20  *
21  *   In either case, Lustre is distributed in the hope that it will be
22  *   useful, but WITHOUT ANY WARRANTY; without even the implied warranty
23  *   of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
24  *   license text for more details.
25  */
26
27 #ifndef EXPORT_SYMTAB
28 # define EXPORT_SYMTAB
29 #endif
30 #define DEBUG_SUBSYSTEM S_LDLM
31
32 #ifdef __KERNEL__
33 # include <libcfs/libcfs.h>
34 #else
35 # include <liblustre.h>
36 #endif
37
38 #include <lustre_dlm.h>
39 #include <obd_class.h>
40 #include <libcfs/list.h>
41 #include "ldlm_internal.h"
42
43 #ifdef __KERNEL__
44 static int ldlm_num_threads;
45 CFS_MODULE_PARM(ldlm_num_threads, "i", int, 0444,
46                 "number of DLM service threads to start");
47 #endif
48
49 extern cfs_mem_cache_t *ldlm_resource_slab;
50 extern cfs_mem_cache_t *ldlm_lock_slab;
51 static struct semaphore ldlm_ref_sem;
52 static int ldlm_refcount;
53
54 /* LDLM state */
55
56 static struct ldlm_state *ldlm_state;
57
58 inline cfs_time_t round_timeout(cfs_time_t timeout)
59 {
60         return cfs_time_seconds((int)cfs_duration_sec(cfs_time_sub(timeout, 0)) + 1);
61 }
62
63 /* timeout for initial callback (AST) reply */
64 static inline unsigned int ldlm_get_rq_timeout(unsigned int ldlm_timeout,
65                                                unsigned int obd_timeout)
66 {
67         unsigned int timeout = min(ldlm_timeout, obd_timeout / 3);
68
69         return timeout < 1 ? 1 : timeout;
70 }
71
72 #ifdef __KERNEL__
73 /* w_l_spinlock protects both waiting_locks_list and expired_lock_thread */
74 static spinlock_t waiting_locks_spinlock;   /* BH lock (timer) */
75 static struct list_head waiting_locks_list;
76 static cfs_timer_t waiting_locks_timer;
77
78 static struct expired_lock_thread {
79         cfs_waitq_t               elt_waitq;
80         int                       elt_state;
81         int                       elt_dump;
82         struct list_head          elt_expired_locks;
83 } expired_lock_thread;
84 #endif
85
86 #define ELT_STOPPED   0
87 #define ELT_READY     1
88 #define ELT_TERMINATE 2
89
90 struct ldlm_bl_pool {
91         spinlock_t              blp_lock;
92
93         /*
94          * blp_prio_list is used for callbacks that should be handled
95          * as a priority. It is used for LDLM_FL_DISCARD_DATA requests.
96          * see bug 13843
97          */
98         struct list_head        blp_prio_list;
99
100         /*
101          * blp_list is used for all other callbacks which are likely
102          * to take longer to process.
103          */
104         struct list_head        blp_list;
105
106         cfs_waitq_t             blp_waitq;
107         struct completion       blp_comp;
108         atomic_t                blp_num_threads;
109         atomic_t                blp_busy_threads;
110         int                     blp_min_threads;
111         int                     blp_max_threads;
112 };
113
114 struct ldlm_bl_work_item {
115         struct list_head        blwi_entry;
116         struct ldlm_namespace   *blwi_ns;
117         struct ldlm_lock_desc   blwi_ld;
118         struct ldlm_lock        *blwi_lock;
119         struct list_head        blwi_head;
120         int                     blwi_count;
121 };
122
123 #ifdef __KERNEL__
124
125 static inline int have_expired_locks(void)
126 {
127         int need_to_run;
128
129         ENTRY;
130         spin_lock_bh(&waiting_locks_spinlock);
131         need_to_run = !list_empty(&expired_lock_thread.elt_expired_locks);
132         spin_unlock_bh(&waiting_locks_spinlock);
133
134         RETURN(need_to_run);
135 }
136
137 static int expired_lock_main(void *arg)
138 {
139         struct list_head *expired = &expired_lock_thread.elt_expired_locks;
140         struct l_wait_info lwi = { 0 };
141         int do_dump;
142
143         ENTRY;
144         cfs_daemonize("ldlm_elt");
145
146         expired_lock_thread.elt_state = ELT_READY;
147         cfs_waitq_signal(&expired_lock_thread.elt_waitq);
148
149         while (1) {
150                 l_wait_event(expired_lock_thread.elt_waitq,
151                              have_expired_locks() ||
152                              expired_lock_thread.elt_state == ELT_TERMINATE,
153                              &lwi);
154
155                 spin_lock_bh(&waiting_locks_spinlock);
156                 if (expired_lock_thread.elt_dump) {
157                         spin_unlock_bh(&waiting_locks_spinlock);
158
159                         /* from waiting_locks_callback, but not in timer */
160                         libcfs_debug_dumplog();
161                         libcfs_run_lbug_upcall(__FILE__,
162                                                 "waiting_locks_callback",
163                                                 expired_lock_thread.elt_dump);
164
165                         spin_lock_bh(&waiting_locks_spinlock);
166                         expired_lock_thread.elt_dump = 0;
167                 }
168
169                 do_dump = 0;
170
171                 while (!list_empty(expired)) {
172                         struct obd_export *export;
173                         struct ldlm_lock *lock;
174
175                         lock = list_entry(expired->next, struct ldlm_lock,
176                                           l_pending_chain);
177                         if ((void *)lock < LP_POISON + CFS_PAGE_SIZE &&
178                             (void *)lock >= LP_POISON) {
179                                 spin_unlock_bh(&waiting_locks_spinlock);
180                                 CERROR("free lock on elt list %p\n", lock);
181                                 LBUG();
182                         }
183                         list_del_init(&lock->l_pending_chain);
184                         if ((void *)lock->l_export < LP_POISON + CFS_PAGE_SIZE &&
185                             (void *)lock->l_export >= LP_POISON) {
186                                 CERROR("lock with free export on elt list %p\n",
187                                        lock->l_export);
188                                 lock->l_export = NULL;
189                                 LDLM_ERROR(lock, "free export");
190                                 continue;
191                         }
192                         export = class_export_get(lock->l_export);
193                         spin_unlock_bh(&waiting_locks_spinlock);
194
195                         do_dump++;
196                         class_fail_export(export);
197                         class_export_put(export);
198                         spin_lock_bh(&waiting_locks_spinlock);
199                 }
200                 spin_unlock_bh(&waiting_locks_spinlock);
201
202                 if (do_dump && obd_dump_on_eviction) {
203                         CERROR("dump the log upon eviction\n");
204                         libcfs_debug_dumplog();
205                 }
206
207                 if (expired_lock_thread.elt_state == ELT_TERMINATE)
208                         break;
209         }
210
211         expired_lock_thread.elt_state = ELT_STOPPED;
212         cfs_waitq_signal(&expired_lock_thread.elt_waitq);
213         RETURN(0);
214 }
215
216 static int ldlm_add_waiting_lock(struct ldlm_lock *lock);
217
218 /* This is called from within a timer interrupt and cannot schedule */
219 static void waiting_locks_callback(unsigned long unused)
220 {
221         struct ldlm_lock *lock, *last = NULL;
222
223 repeat:
224         spin_lock_bh(&waiting_locks_spinlock);
225         while (!list_empty(&waiting_locks_list)) {
226                 lock = list_entry(waiting_locks_list.next, struct ldlm_lock,
227                                   l_pending_chain);
228
229                 if (cfs_time_after(lock->l_callback_timeout, cfs_time_current()) ||
230                     (lock->l_req_mode == LCK_GROUP))
231                         break;
232
233                 if (ptlrpc_check_suspend()) {
234                         /* there is a case when we talk to one mds, holding
235                          * lock from another mds. this way we easily can get
236                          * here, if second mds is being recovered. so, we
237                          * suspend timeouts. bug 6019 */
238
239                         LDLM_ERROR(lock, "recharge timeout: %s@%s nid %s ",
240                                    lock->l_export->exp_client_uuid.uuid,
241                                    lock->l_export->exp_connection->c_remote_uuid.uuid,
242                                    libcfs_nid2str(lock->l_export->exp_connection->c_peer.nid));
243
244                         list_del_init(&lock->l_pending_chain);
245                         spin_unlock_bh(&waiting_locks_spinlock);
246                         ldlm_add_waiting_lock(lock);
247                         goto repeat;
248                 }
249
250                 /* if timeout overlaps the activation time of suspended timeouts
251                  * then extend it to give a chance for client to reconnect */
252                 if (cfs_time_before(cfs_time_sub(lock->l_callback_timeout,
253                                                  cfs_time_seconds(obd_timeout)/2),
254                                     ptlrpc_suspend_wakeup_time())) {
255                         LDLM_ERROR(lock, "extend timeout due to recovery: %s@%s nid %s ",
256                                    lock->l_export->exp_client_uuid.uuid,
257                                    lock->l_export->exp_connection->c_remote_uuid.uuid,
258                                    libcfs_nid2str(lock->l_export->exp_connection->c_peer.nid));
259
260                         list_del_init(&lock->l_pending_chain);
261                         spin_unlock_bh(&waiting_locks_spinlock);
262                         ldlm_add_waiting_lock(lock);
263                         goto repeat;
264                 }
265
266                 LDLM_ERROR(lock, "lock callback timer expired: evicting client "
267                            "%s@%s nid %s\n",
268                            lock->l_export->exp_client_uuid.uuid,
269                            lock->l_export->exp_connection->c_remote_uuid.uuid,
270                            libcfs_nid2str(lock->l_export->exp_connection->c_peer.nid));
271
272                 last = lock;
273
274                 list_del(&lock->l_pending_chain);
275                 list_add(&lock->l_pending_chain,
276                          &expired_lock_thread.elt_expired_locks);
277         }
278
279         if (!list_empty(&expired_lock_thread.elt_expired_locks)) {
280                 if (obd_dump_on_timeout)
281                         expired_lock_thread.elt_dump = __LINE__;
282
283                 cfs_waitq_signal(&expired_lock_thread.elt_waitq);
284         }
285
286         /*
287          * Make sure the timer will fire again if we have any locks
288          * left.
289          */
290         if (!list_empty(&waiting_locks_list)) {
291                 cfs_time_t timeout_rounded;
292                 lock = list_entry(waiting_locks_list.next, struct ldlm_lock,
293                                   l_pending_chain);
294                 timeout_rounded = (cfs_time_t)round_timeout(lock->l_callback_timeout);
295                 cfs_timer_arm(&waiting_locks_timer, timeout_rounded);
296         }
297         spin_unlock_bh(&waiting_locks_spinlock);
298 }
299
300 /*
301  * Indicate that we're waiting for a client to call us back cancelling a given
302  * lock.  We add it to the pending-callback chain, and schedule the lock-timeout
303  * timer to fire appropriately.  (We round up to the next second, to avoid
304  * floods of timer firings during periods of high lock contention and traffic).
305  *
306  * Called with the namespace lock held.
307  */
308 static int __ldlm_add_waiting_lock(struct ldlm_lock *lock)
309 {
310         cfs_time_t timeout_rounded;
311
312         if (!list_empty(&lock->l_pending_chain))
313                 return 0;
314
315         lock->l_callback_timeout =cfs_time_add(cfs_time_current(),
316                                                cfs_time_seconds(obd_timeout)/2);
317
318         timeout_rounded = round_timeout(lock->l_callback_timeout);
319
320         if (cfs_time_before(timeout_rounded, cfs_timer_deadline(&waiting_locks_timer)) ||
321             !cfs_timer_is_armed(&waiting_locks_timer)) {
322                 cfs_timer_arm(&waiting_locks_timer, timeout_rounded);
323
324         }
325         list_add_tail(&lock->l_pending_chain, &waiting_locks_list); /* FIFO */
326         return 1;
327 }
328
329 static int ldlm_add_waiting_lock(struct ldlm_lock *lock)
330 {
331         int ret;
332
333         LASSERT(!(lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK));
334
335         spin_lock_bh(&waiting_locks_spinlock);
336         if (lock->l_destroyed) {
337                 static cfs_time_t next;
338                 spin_unlock_bh(&waiting_locks_spinlock);
339                 LDLM_ERROR(lock, "not waiting on destroyed lock (bug 5653)");
340                 if (cfs_time_after(cfs_time_current(), next)) {
341                         next = cfs_time_shift(14400);
342                         libcfs_debug_dumpstack(NULL);
343                 }
344                 return 0;
345         }
346
347         ret = __ldlm_add_waiting_lock(lock);
348         spin_unlock_bh(&waiting_locks_spinlock);
349
350         LDLM_DEBUG(lock, "%sadding to wait list",
351                    ret == 0 ? "not re-" : "");
352         return ret;
353 }
354
355 /*
356  * Remove a lock from the pending list, likely because it had its cancellation
357  * callback arrive without incident.  This adjusts the lock-timeout timer if
358  * needed.  Returns 0 if the lock wasn't pending after all, 1 if it was.
359  *
360  * Called with namespace lock held.
361  */
362 int __ldlm_del_waiting_lock(struct ldlm_lock *lock)
363 {
364         struct list_head *list_next;
365
366         if (list_empty(&lock->l_pending_chain))
367                 return 0;
368
369         list_next = lock->l_pending_chain.next;
370         if (lock->l_pending_chain.prev == &waiting_locks_list) {
371                 /* Removing the head of the list, adjust timer. */
372                 if (list_next == &waiting_locks_list) {
373                         /* No more, just cancel. */
374                         cfs_timer_disarm(&waiting_locks_timer);
375                 } else {
376                         struct ldlm_lock *next;
377                         next = list_entry(list_next, struct ldlm_lock,
378                                           l_pending_chain);
379                         cfs_timer_arm(&waiting_locks_timer,
380                                       round_timeout(next->l_callback_timeout));
381                 }
382         }
383         list_del_init(&lock->l_pending_chain);
384
385         return 1;
386 }
387
388 int ldlm_del_waiting_lock(struct ldlm_lock *lock)
389 {
390         int ret;
391
392         if (lock->l_export == NULL) {
393                 /* We don't have a "waiting locks list" on clients. */
394                 LDLM_DEBUG(lock, "client lock: no-op");
395                 return 0;
396         }
397
398         spin_lock_bh(&waiting_locks_spinlock);
399         ret = __ldlm_del_waiting_lock(lock);
400         spin_unlock_bh(&waiting_locks_spinlock);
401
402         LDLM_DEBUG(lock, "%s", ret == 0 ? "wasn't waiting" : "removed");
403         return ret;
404 }
405
406 /*
407  * Prolong the lock
408  *
409  * Called with namespace lock held.
410  */
411 int ldlm_refresh_waiting_lock(struct ldlm_lock *lock)
412 {
413         if (lock->l_export == NULL) {
414                 /* We don't have a "waiting locks list" on clients. */
415                 LDLM_DEBUG(lock, "client lock: no-op");
416                 return 0;
417         }
418
419         spin_lock_bh(&waiting_locks_spinlock);
420
421         if (list_empty(&lock->l_pending_chain)) {
422                 spin_unlock_bh(&waiting_locks_spinlock);
423                 LDLM_DEBUG(lock, "wasn't waiting");
424                 return 0;
425         }
426
427         __ldlm_del_waiting_lock(lock);
428         __ldlm_add_waiting_lock(lock);
429         spin_unlock_bh(&waiting_locks_spinlock);
430
431         LDLM_DEBUG(lock, "refreshed");
432         return 1;
433 }
434
435 #else /* !__KERNEL__ */
436
437 static int ldlm_add_waiting_lock(struct ldlm_lock *lock)
438 {
439         LASSERT(!(lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK));
440         RETURN(1);
441 }
442
443 int ldlm_del_waiting_lock(struct ldlm_lock *lock)
444 {
445         RETURN(0);
446 }
447
448 int ldlm_refresh_waiting_lock(struct ldlm_lock *lock)
449 {
450         RETURN(0);
451 }
452 #endif /* __KERNEL__ */
453
454 static void ldlm_failed_ast(struct ldlm_lock *lock, int rc,
455                             const char *ast_type)
456 {
457         struct ptlrpc_connection *conn = lock->l_export->exp_connection;
458         char                     *str = libcfs_nid2str(conn->c_peer.nid);
459
460         LCONSOLE_ERROR_MSG(0x138, "A client on nid %s was evicted from "
461                            "service %s.\n", str,
462                            lock->l_export->exp_obd->obd_name);
463
464         LCONSOLE_ERROR_MSG(0x012, "Lock %s callback to %s timed out for "
465                            "resource %d\n", ast_type,
466                            obd_export_nid2str(lock->l_export), rc);
467
468         if (obd_dump_on_timeout)
469                 libcfs_debug_dumplog();
470         class_fail_export(lock->l_export);
471 }
472
473 static int ldlm_handle_ast_error(struct ldlm_lock *lock,
474                                  struct ptlrpc_request *req, int rc,
475                                  const char *ast_type)
476 {
477         lnet_process_id_t peer = req->rq_import->imp_connection->c_peer;
478
479         if (rc == -ETIMEDOUT || rc == -EINTR || rc == -ENOTCONN) {
480                 LASSERT(lock->l_export);
481                 if (lock->l_export->exp_libclient) {
482                         LDLM_DEBUG(lock, "%s AST to liblustre client (nid %s)"
483                                    " timeout, just cancelling lock", ast_type,
484                                    libcfs_nid2str(peer.nid));
485                         ldlm_lock_cancel(lock);
486                         rc = -ERESTART;
487                 } else if (lock->l_flags & LDLM_FL_CANCEL) {
488                         LDLM_DEBUG(lock, "%s AST timeout from nid %s, but "
489                                    "cancel was received (AST reply lost?)",
490                                    ast_type, libcfs_nid2str(peer.nid));
491                         ldlm_lock_cancel(lock);
492                         rc = -ERESTART;
493                 } else {
494                         ldlm_del_waiting_lock(lock);
495                         ldlm_failed_ast(lock, rc, ast_type);
496                 }
497         } else if (rc) {
498                 if (rc == -EINVAL)
499                         LDLM_DEBUG(lock, "client (nid %s) returned %d"
500                                " from %s AST - normal race",
501                                libcfs_nid2str(peer.nid),
502                                req->rq_repmsg ?
503                                lustre_msg_get_status(req->rq_repmsg) : -1,
504                                ast_type);
505                 else
506                         LDLM_ERROR(lock, "client (nid %s) returned %d "
507                                    "from %s AST", libcfs_nid2str(peer.nid),
508                                    (req->rq_repmsg != NULL) ?
509                                    lustre_msg_get_status(req->rq_repmsg) : 0,
510                                    ast_type);
511                 ldlm_lock_cancel(lock);
512                 /* Server-side AST functions are called from ldlm_reprocess_all,
513                  * which needs to be told to please restart its reprocessing. */
514                 rc = -ERESTART;
515         }
516
517         return rc;
518 }
519
520 static int ldlm_cb_interpret(struct ptlrpc_request *req, void *data, int rc)
521 {
522         struct ldlm_cb_set_arg *arg;
523         struct ldlm_lock *lock;
524         ENTRY;
525
526         LASSERT(data != NULL);
527
528         arg = req->rq_async_args.pointer_arg[0];
529         lock = req->rq_async_args.pointer_arg[1];
530         LASSERT(lock != NULL);
531         if (rc != 0) {
532                 /* If client canceled the lock but the cancel has not 
533                  * been recieved yet, we need to update lvbo to have the
534                  * proper attributes cached. */
535                 if (rc == -EINVAL && arg->type == LDLM_BL_CALLBACK)
536                         ldlm_res_lvbo_update(lock->l_resource, NULL, 
537                                              0, 1);
538                 rc = ldlm_handle_ast_error(lock, req, rc, 
539                                            arg->type == LDLM_BL_CALLBACK
540                                            ? "blocking" : "completion");
541         }                
542
543         LDLM_LOCK_PUT(lock);
544
545         if (rc == -ERESTART)
546                 atomic_set(&arg->restart, 1);
547
548         RETURN(0);
549 }
550
551 static inline int ldlm_bl_and_cp_ast_fini(struct ptlrpc_request *req,
552                                           struct ldlm_cb_set_arg *arg,
553                                           struct ldlm_lock *lock,
554                                           int instant_cancel)
555 {
556         int rc = 0;
557         ENTRY;
558
559         if (unlikely(instant_cancel)) {
560                 rc = ptl_send_rpc(req, 1);
561                 ptlrpc_req_finished(req);
562                 if (rc == 0)
563                         /* If we cancelled the lock, we need to restart
564                          * ldlm_reprocess_queue */
565                         atomic_set(&arg->restart, 1);
566         } else {
567                 LDLM_LOCK_GET(lock);
568                 ptlrpc_set_add_req(arg->set, req);
569         }       
570
571         RETURN(rc);
572 }
573
574 /*
575  * ->l_blocking_ast() method for server-side locks. This is invoked when newly
576  * enqueued server lock conflicts with given one.
577  *
578  * Sends blocking ast rpc to the client owning that lock; arms timeout timer
579  * to wait for client response.
580  */
581 int ldlm_server_blocking_ast(struct ldlm_lock *lock,
582                              struct ldlm_lock_desc *desc,
583                              void *data, int flag)
584 {
585         struct ldlm_cb_set_arg *arg = data;
586         struct ldlm_request *body;
587         struct ptlrpc_request *req;
588         int size[] = { [MSG_PTLRPC_BODY_OFF] = sizeof(struct ptlrpc_body),
589                        [DLM_LOCKREQ_OFF]     = sizeof(*body) };
590         int instant_cancel = 0, rc;
591         ENTRY;
592
593         if (flag == LDLM_CB_CANCELING) {
594                 /* Don't need to do anything here. */
595                 RETURN(0);
596         }
597
598         LASSERT(lock);
599         LASSERT(data != NULL);
600         if (lock->l_export->exp_obd->obd_recovering != 0) {
601                 LDLM_ERROR(lock, "BUG 6063: lock collide during recovery");
602                 ldlm_lock_dump(D_ERROR, lock, 0);
603         }
604
605         req = ptlrpc_prep_req(lock->l_export->exp_imp_reverse,
606                               LUSTRE_DLM_VERSION, LDLM_BL_CALLBACK, 2, size,
607                               NULL);
608         if (req == NULL)
609                 RETURN(-ENOMEM);
610
611         req->rq_async_args.pointer_arg[0] = arg;
612         req->rq_async_args.pointer_arg[1] = lock;
613         req->rq_interpret_reply = ldlm_cb_interpret;
614         req->rq_no_resend = 1;
615
616         lock_res(lock->l_resource);
617         if (lock->l_granted_mode != lock->l_req_mode) {
618                 /* this blocking AST will be communicated as part of the
619                  * completion AST instead */
620                 unlock_res(lock->l_resource);
621                 ptlrpc_req_finished(req);
622                 LDLM_DEBUG(lock, "lock not granted, not sending blocking AST");
623                 RETURN(0);
624         }
625
626         if (lock->l_destroyed) {
627                 /* What's the point? */
628                 unlock_res(lock->l_resource);
629                 ptlrpc_req_finished(req);
630                 RETURN(0);
631         }
632
633         if (lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK)
634                 instant_cancel = 1;
635
636         body = lustre_msg_buf(req->rq_reqmsg, DLM_LOCKREQ_OFF, sizeof(*body));
637         body->lock_handle[0] = lock->l_remote_handle;
638         body->lock_desc = *desc;
639         body->lock_flags |= (lock->l_flags & LDLM_AST_FLAGS);
640
641         LDLM_DEBUG(lock, "server preparing blocking AST");
642
643         ptlrpc_req_set_repsize(req, 1, NULL);
644         if (instant_cancel) {
645                 unlock_res(lock->l_resource);
646                 ldlm_lock_cancel(lock);
647         } else {
648                 LASSERT(lock->l_granted_mode == lock->l_req_mode);
649                 ldlm_add_waiting_lock(lock);
650                 unlock_res(lock->l_resource);
651         }
652
653         req->rq_send_state = LUSTRE_IMP_FULL;
654         req->rq_timeout = ldlm_get_rq_timeout(ldlm_timeout, obd_timeout);
655
656         if (lock->l_export && lock->l_export->exp_ldlm_stats)
657                 lprocfs_counter_incr(lock->l_export->exp_ldlm_stats,
658                                      LDLM_BL_CALLBACK - LDLM_FIRST_OPC);
659
660         rc = ldlm_bl_and_cp_ast_fini(req, arg, lock, instant_cancel);
661
662         RETURN(rc);
663 }
664
665 int ldlm_server_completion_ast(struct ldlm_lock *lock, int flags, void *data)
666 {
667         struct ldlm_cb_set_arg *arg = data;
668         struct ldlm_request *body;
669         struct ptlrpc_request *req;
670         struct timeval granted_time;
671         long total_enqueue_wait;
672         int size[3] = { [MSG_PTLRPC_BODY_OFF] = sizeof(struct ptlrpc_body),
673                         [DLM_LOCKREQ_OFF]     = sizeof(*body) };
674         int rc, buffers = 2, instant_cancel = 0;
675         ENTRY;
676
677         LASSERT(lock != NULL);
678         LASSERT(data != NULL);
679
680         do_gettimeofday(&granted_time);
681         total_enqueue_wait = cfs_timeval_sub(&granted_time,
682                                              &lock->l_enqueued_time, NULL);
683
684         if (total_enqueue_wait / 1000000 > obd_timeout)
685                 LDLM_ERROR(lock, "enqueue wait took %luus from %lu",
686                            total_enqueue_wait, lock->l_enqueued_time.tv_sec);
687
688         lock_res_and_lock(lock);
689         if (lock->l_resource->lr_lvb_len) {
690                 size[DLM_REQ_REC_OFF] = lock->l_resource->lr_lvb_len;
691                 buffers = 3;
692         }
693         unlock_res_and_lock(lock);
694
695         req = ptlrpc_prep_req(lock->l_export->exp_imp_reverse,
696                               LUSTRE_DLM_VERSION, LDLM_CP_CALLBACK, buffers,
697                               size, NULL);
698         if (req == NULL)
699                 RETURN(-ENOMEM);
700
701         req->rq_async_args.pointer_arg[0] = arg;
702         req->rq_async_args.pointer_arg[1] = lock;
703         req->rq_interpret_reply = ldlm_cb_interpret;
704         req->rq_no_resend = 1;
705
706         body = lustre_msg_buf(req->rq_reqmsg, DLM_LOCKREQ_OFF, sizeof(*body));
707         body->lock_handle[0] = lock->l_remote_handle;
708         body->lock_flags = flags;
709         ldlm_lock2desc(lock, &body->lock_desc);
710
711         if (buffers == 3) {
712                 void *lvb;
713
714                 lvb = lustre_msg_buf(req->rq_reqmsg, DLM_REQ_REC_OFF,
715                                      lock->l_resource->lr_lvb_len);
716                 lock_res_and_lock(lock);
717                 memcpy(lvb, lock->l_resource->lr_lvb_data,
718                        lock->l_resource->lr_lvb_len);
719                 unlock_res_and_lock(lock);
720         }
721
722         LDLM_DEBUG(lock, "server preparing completion AST (after %ldus wait)",
723                    total_enqueue_wait);
724
725         ptlrpc_req_set_repsize(req, 1, NULL);
726
727         req->rq_send_state = LUSTRE_IMP_FULL;
728         req->rq_timeout = ldlm_get_rq_timeout(ldlm_timeout, obd_timeout);
729
730         /* We only send real blocking ASTs after the lock is granted */
731         lock_res_and_lock(lock);
732         if (lock->l_flags & LDLM_FL_AST_SENT) {
733                 body->lock_flags |= LDLM_FL_AST_SENT;
734
735                 /* We might get here prior to ldlm_handle_enqueue setting
736                  * LDLM_FL_CANCEL_ON_BLOCK flag. Then we will put this lock
737                  * into waiting list, but this is safe and similar code in
738                  * ldlm_handle_enqueue will call ldlm_lock_cancel() still,
739                  * that would not only cancel the lock, but will also remove
740                  * it from waiting list */
741                 if (lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK) {
742                         unlock_res_and_lock(lock);
743                         ldlm_lock_cancel(lock);
744                         instant_cancel = 1;
745                         lock_res_and_lock(lock);
746                 } else {
747                         /* start the lock-timeout clock */
748                         ldlm_add_waiting_lock(lock);
749                 }
750         }
751         unlock_res_and_lock(lock);
752
753         if (lock->l_export && lock->l_export->exp_ldlm_stats)
754                 lprocfs_counter_incr(lock->l_export->exp_ldlm_stats,
755                                      LDLM_CP_CALLBACK - LDLM_FIRST_OPC);
756
757         rc = ldlm_bl_and_cp_ast_fini(req, arg, lock, instant_cancel);
758
759         RETURN(rc);
760 }
761
762 int ldlm_server_glimpse_ast(struct ldlm_lock *lock, void *data)
763 {
764         struct ldlm_resource *res = lock->l_resource;
765         struct ldlm_request *body;
766         struct ptlrpc_request *req;
767         int size[] = { [MSG_PTLRPC_BODY_OFF] = sizeof(struct ptlrpc_body),
768                        [DLM_LOCKREQ_OFF]     = sizeof(*body) };
769         int rc = 0;
770         ENTRY;
771
772         LASSERT(lock != NULL);
773
774         req = ptlrpc_prep_req(lock->l_export->exp_imp_reverse,
775                               LUSTRE_DLM_VERSION, LDLM_GL_CALLBACK, 2, size,
776                               NULL);
777         if (req == NULL)
778                 RETURN(-ENOMEM);
779
780         body = lustre_msg_buf(req->rq_reqmsg, DLM_LOCKREQ_OFF, sizeof(*body));
781         body->lock_handle[0] = lock->l_remote_handle;
782         ldlm_lock2desc(lock, &body->lock_desc);
783
784         lock_res_and_lock(lock);
785         size[REPLY_REC_OFF] = lock->l_resource->lr_lvb_len;
786         unlock_res_and_lock(lock);
787         res = lock->l_resource;
788         ptlrpc_req_set_repsize(req, 2, size);
789
790         req->rq_send_state = LUSTRE_IMP_FULL;
791         req->rq_timeout = ldlm_get_rq_timeout(ldlm_timeout, obd_timeout);
792
793         if (lock->l_export && lock->l_export->exp_ldlm_stats)
794                 lprocfs_counter_incr(lock->l_export->exp_ldlm_stats,
795                                      LDLM_GL_CALLBACK - LDLM_FIRST_OPC);
796
797         rc = ptlrpc_queue_wait(req);
798         if (rc == -ELDLM_NO_LOCK_DATA)
799                 LDLM_DEBUG(lock, "lost race - client has a lock but no inode");
800         else if (rc != 0)
801                 rc = ldlm_handle_ast_error(lock, req, rc, "glimpse");
802         else
803                 rc = ldlm_res_lvbo_update(res, req->rq_repmsg,
804                                           REPLY_REC_OFF, 1);
805         ptlrpc_req_finished(req);
806         RETURN(rc);
807 }
808
809 static struct ldlm_lock *
810 find_existing_lock(struct obd_export *exp,
811                    const struct lustre_handle *remote_hdl)
812 {
813         struct list_head *iter;
814
815         spin_lock(&exp->exp_ldlm_data.led_lock);
816         list_for_each(iter, &exp->exp_ldlm_data.led_held_locks) {
817                 struct ldlm_lock *lock;
818                 lock = list_entry(iter, struct ldlm_lock, l_export_chain);
819                 if (lock->l_remote_handle.cookie == remote_hdl->cookie) {
820                         LDLM_LOCK_GET(lock);
821                         spin_unlock(&exp->exp_ldlm_data.led_lock);
822                         return lock;
823                 }
824         }
825         spin_unlock(&exp->exp_ldlm_data.led_lock);
826         return NULL;
827 }
828
829 #ifdef __KERNEL__
830 extern unsigned long long lu_time_stamp_get(void);
831 #else
832 #define lu_time_stamp_get() time(NULL)
833 #endif
834
835 /*
836  * Main server-side entry point into LDLM. This is called by ptlrpc service
837  * threads to carry out client lock enqueueing requests.
838  */
839 int ldlm_handle_enqueue0(struct ldlm_namespace *ns,
840                          struct ptlrpc_request *req,
841                          const struct ldlm_request *dlm_req,
842                          const struct ldlm_callback_suite *cbs)
843 {
844         struct ldlm_reply *dlm_rep;
845         int size[3] = { [MSG_PTLRPC_BODY_OFF] = sizeof(struct ptlrpc_body),
846                         [DLM_LOCKREPLY_OFF]   = sizeof(*dlm_rep) };
847         int rc = 0;
848         __u32 flags;
849         ldlm_error_t err = ELDLM_OK;
850         struct ldlm_lock *lock = NULL;
851         void *cookie = NULL;
852         ENTRY;
853
854         LDLM_DEBUG_NOLOCK("server-side enqueue handler START");
855
856         ldlm_request_cancel(req, dlm_req, LDLM_ENQUEUE_CANCEL_OFF);
857         flags = dlm_req->lock_flags;
858
859         LASSERT(req->rq_export);
860
861         if (req->rq_export->exp_ldlm_stats)
862                 lprocfs_counter_incr(req->rq_export->exp_ldlm_stats,
863                                      LDLM_ENQUEUE - LDLM_FIRST_OPC);
864
865         if (unlikely(dlm_req->lock_desc.l_resource.lr_type < LDLM_MIN_TYPE ||
866                      dlm_req->lock_desc.l_resource.lr_type >= LDLM_MAX_TYPE)) {
867                 DEBUG_REQ(D_ERROR, req, "invalid lock request type %d",
868                           dlm_req->lock_desc.l_resource.lr_type);
869                 GOTO(out, rc = -EFAULT);
870         }
871
872         if (unlikely(dlm_req->lock_desc.l_req_mode <= LCK_MINMODE ||
873                      dlm_req->lock_desc.l_req_mode >= LCK_MAXMODE ||
874                      dlm_req->lock_desc.l_req_mode &
875                      (dlm_req->lock_desc.l_req_mode-1))) {
876                 DEBUG_REQ(D_ERROR, req, "invalid lock request mode %d",
877                           dlm_req->lock_desc.l_req_mode);
878                 GOTO(out, rc = -EFAULT);
879         }
880
881         if (req->rq_export->exp_connect_flags & OBD_CONNECT_IBITS) {
882                 if (unlikely(dlm_req->lock_desc.l_resource.lr_type ==
883                              LDLM_PLAIN)) {
884                         DEBUG_REQ(D_ERROR, req,
885                                   "PLAIN lock request from IBITS client?");
886                         GOTO(out, rc = -EPROTO);
887                 }
888         } else if (unlikely(dlm_req->lock_desc.l_resource.lr_type ==
889                             LDLM_IBITS)) {
890                 DEBUG_REQ(D_ERROR, req,
891                           "IBITS lock request from unaware client?");
892                 GOTO(out, rc = -EPROTO);
893         }
894
895 #if 0
896         /* FIXME this makes it impossible to use LDLM_PLAIN locks -- check
897            against server's _CONNECT_SUPPORTED flags? (I don't want to use
898            ibits for mgc/mgs) */
899
900         /* INODEBITS_INTEROP: Perform conversion from plain lock to
901          * inodebits lock if client does not support them. */
902         if (!(req->rq_export->exp_connect_flags & OBD_CONNECT_IBITS) &&
903             (dlm_req->lock_desc.l_resource.lr_type == LDLM_PLAIN)) {
904                 dlm_req->lock_desc.l_resource.lr_type = LDLM_IBITS;
905                 dlm_req->lock_desc.l_policy_data.l_inodebits.bits =
906                         MDS_INODELOCK_LOOKUP | MDS_INODELOCK_UPDATE;
907                 if (dlm_req->lock_desc.l_req_mode == LCK_PR)
908                         dlm_req->lock_desc.l_req_mode = LCK_CR;
909         }
910 #endif
911
912         if (unlikely(flags & LDLM_FL_REPLAY)) {
913                 lock = find_existing_lock(req->rq_export,
914                                           &dlm_req->lock_handle[0]);
915                 if (lock != NULL) {
916                         DEBUG_REQ(D_DLMTRACE, req, "found existing lock cookie "
917                                   LPX64, lock->l_handle.h_cookie);
918                         GOTO(existing_lock, rc = 0);
919                 }
920         }
921
922         /* The lock's callback data might be set in the policy function */
923         lock = ldlm_lock_create(ns, &dlm_req->lock_desc.l_resource.lr_name,
924                                 dlm_req->lock_desc.l_resource.lr_type,
925                                 dlm_req->lock_desc.l_req_mode,
926                                 cbs->lcs_blocking, cbs->lcs_completion,
927                                 cbs->lcs_glimpse, NULL, 0);
928
929         if (!lock)
930                 GOTO(out, rc = -ENOMEM);
931
932         do_gettimeofday(&lock->l_enqueued_time);
933         lock->l_remote_handle = dlm_req->lock_handle[0];
934         LDLM_DEBUG(lock, "server-side enqueue handler, new lock created");
935
936         OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_ENQUEUE_BLOCKED, obd_timeout * 2);
937         /* Don't enqueue a lock onto the export if it has already
938          * been evicted.  Cancel it now instead. (bug 3822) */
939         if (req->rq_export->exp_failed) {
940                 LDLM_ERROR(lock, "lock on destroyed export %p", req->rq_export);
941                 GOTO(out, rc = -ENOTCONN);
942         }
943         lock->l_export = class_export_get(req->rq_export);
944         spin_lock(&lock->l_export->exp_ldlm_data.led_lock);
945         list_add(&lock->l_export_chain,
946                  &lock->l_export->exp_ldlm_data.led_held_locks);
947         spin_unlock(&lock->l_export->exp_ldlm_data.led_lock);
948
949 existing_lock:
950
951         if (flags & LDLM_FL_HAS_INTENT) {
952                 /* In this case, the reply buffer is allocated deep in
953                  * local_lock_enqueue by the policy function. */
954                 cookie = req;
955         } else {
956                 int buffers = 2;
957
958                 lock_res_and_lock(lock);
959                 if (lock->l_resource->lr_lvb_len) {
960                         size[DLM_REPLY_REC_OFF] = lock->l_resource->lr_lvb_len;
961                         buffers = 3;
962                 }
963                 unlock_res_and_lock(lock);
964
965                 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_ENQUEUE_EXTENT_ERR))
966                         GOTO(out, rc = -ENOMEM);
967
968                 rc = lustre_pack_reply(req, buffers, size, NULL);
969                 if (rc)
970                         GOTO(out, rc);
971         }
972
973         if (dlm_req->lock_desc.l_resource.lr_type != LDLM_PLAIN)
974                 lock->l_policy_data = dlm_req->lock_desc.l_policy_data;
975         if (dlm_req->lock_desc.l_resource.lr_type == LDLM_EXTENT)
976                 lock->l_req_extent = lock->l_policy_data.l_extent;
977
978         err = ldlm_lock_enqueue(ns, &lock, cookie, (int *)&flags);
979         if (err)
980                 GOTO(out, err);
981
982         dlm_rep = lustre_msg_buf(req->rq_repmsg, DLM_LOCKREPLY_OFF,
983                                  sizeof(*dlm_rep));
984         dlm_rep->lock_flags = flags;
985
986         ldlm_lock2desc(lock, &dlm_rep->lock_desc);
987         ldlm_lock2handle(lock, &dlm_rep->lock_handle);
988
989         /* We never send a blocking AST until the lock is granted, but
990          * we can tell it right now */
991         lock_res_and_lock(lock);
992
993         /* Now take into account flags to be inherited from original lock
994            request both in reply to client and in our own lock flags. */
995         dlm_rep->lock_flags |= dlm_req->lock_flags & LDLM_INHERIT_FLAGS;
996         lock->l_flags |= dlm_req->lock_flags & LDLM_INHERIT_FLAGS;
997
998         /* Don't move a pending lock onto the export if it has already
999          * been evicted.  Cancel it now instead. (bug 5683) */
1000         if (unlikely(req->rq_export->exp_failed ||
1001                      OBD_FAIL_CHECK(OBD_FAIL_LDLM_ENQUEUE_OLD_EXPORT))) {
1002                 LDLM_ERROR(lock, "lock on destroyed export %p", req->rq_export);
1003                 rc = -ENOTCONN;
1004         } else if (lock->l_flags & LDLM_FL_AST_SENT) {
1005                 dlm_rep->lock_flags |= LDLM_FL_AST_SENT;
1006                 if (lock->l_granted_mode == lock->l_req_mode) {
1007                         /*
1008                          * Only cancel lock if it was granted, because it would
1009                          * be destroyed immediatelly and would never be granted
1010                          * in the future, causing timeouts on client.  Not
1011                          * granted lock will be cancelled immediatelly after
1012                          * sending completion AST.
1013                          */
1014                         if (dlm_rep->lock_flags & LDLM_FL_CANCEL_ON_BLOCK) {
1015                                 unlock_res_and_lock(lock);
1016                                 ldlm_lock_cancel(lock);
1017                                 lock_res_and_lock(lock);
1018                         } else
1019                                 ldlm_add_waiting_lock(lock);
1020                 }
1021         }
1022         /* Make sure we never ever grant usual metadata locks to liblustre
1023            clients */
1024         if ((dlm_req->lock_desc.l_resource.lr_type == LDLM_PLAIN ||
1025             dlm_req->lock_desc.l_resource.lr_type == LDLM_IBITS) &&
1026              req->rq_export->exp_libclient) {
1027                 if (unlikely(!(lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK) ||
1028                              !(dlm_rep->lock_flags & LDLM_FL_CANCEL_ON_BLOCK))){
1029                         CERROR("Granting sync lock to libclient. "
1030                                "req fl %d, rep fl %d, lock fl %d\n",
1031                                dlm_req->lock_flags, dlm_rep->lock_flags,
1032                                lock->l_flags);
1033                         LDLM_ERROR(lock, "sync lock");
1034                         if (dlm_req->lock_flags & LDLM_FL_HAS_INTENT) {
1035                                 struct ldlm_intent *it;
1036                                 it = lustre_msg_buf(req->rq_reqmsg,
1037                                                     DLM_INTENT_IT_OFF,
1038                                                     sizeof(*it));
1039                                 if (it != NULL) {
1040                                         CERROR("This is intent %s ("LPU64")\n",
1041                                                ldlm_it2str(it->opc), it->opc);
1042                                 }
1043                         }
1044                 }
1045         }
1046
1047         unlock_res_and_lock(lock);
1048
1049         EXIT;
1050  out:
1051         req->rq_status = err;
1052         if (req->rq_reply_state == NULL) {
1053                 err = lustre_pack_reply(req, 1, NULL, NULL);
1054                 if (rc == 0)
1055                         rc = err;
1056                 req->rq_status = rc;
1057         }
1058
1059         /* The LOCK_CHANGED code in ldlm_lock_enqueue depends on this
1060          * ldlm_reprocess_all.  If this moves, revisit that code. -phil */
1061         if (lock) {
1062                 LDLM_DEBUG(lock, "server-side enqueue handler, sending reply"
1063                            "(err=%d, rc=%d)", err, rc);
1064
1065                 lock_res_and_lock(lock);
1066                 if (rc == 0) {
1067                         size[DLM_REPLY_REC_OFF] = lock->l_resource->lr_lvb_len;
1068                         if (size[DLM_REPLY_REC_OFF] > 0) {
1069                                 void *lvb = lustre_msg_buf(req->rq_repmsg,
1070                                                        DLM_REPLY_REC_OFF,
1071                                                        size[DLM_REPLY_REC_OFF]);
1072                                 LASSERTF(lvb != NULL, "req %p, lock %p\n",
1073                                          req, lock);
1074
1075                                 memcpy(lvb, lock->l_resource->lr_lvb_data,
1076                                        size[DLM_REPLY_REC_OFF]);
1077                         }
1078                 } else {
1079                         ldlm_resource_unlink_lock(lock);
1080                         ldlm_lock_destroy_nolock(lock);
1081                 }
1082                 unlock_res_and_lock(lock);
1083
1084                 if (!err && dlm_req->lock_desc.l_resource.lr_type != LDLM_FLOCK)
1085                         ldlm_reprocess_all(lock->l_resource);
1086
1087                 LDLM_LOCK_PUT(lock);
1088         }
1089
1090         LDLM_DEBUG_NOLOCK("server-side enqueue handler END (lock %p, rc %d)",
1091                           lock, rc);
1092
1093         return rc;
1094 }
1095
1096 int ldlm_handle_enqueue(struct ptlrpc_request *req,
1097                         ldlm_completion_callback completion_callback,
1098                         ldlm_blocking_callback blocking_callback,
1099                         ldlm_glimpse_callback glimpse_callback)
1100 {
1101         int rc;
1102         struct ldlm_request *dlm_req;
1103         struct ldlm_callback_suite cbs = {
1104                 .lcs_completion = completion_callback,
1105                 .lcs_blocking   = blocking_callback,
1106                 .lcs_glimpse    = glimpse_callback
1107         };
1108
1109
1110         dlm_req = lustre_swab_reqbuf(req, DLM_LOCKREQ_OFF,
1111                                      sizeof *dlm_req, lustre_swab_ldlm_request);
1112         if (dlm_req != NULL) {
1113                 rc = ldlm_handle_enqueue0(req->rq_export->exp_obd->obd_namespace,
1114                                           req, dlm_req, &cbs);
1115         } else {
1116                 CERROR ("Can't unpack dlm_req\n");
1117                 rc = -EFAULT;
1118         }
1119         return rc;
1120 }
1121
1122 int ldlm_handle_convert0(struct ptlrpc_request *req,
1123                          const struct ldlm_request *dlm_req)
1124 {
1125         struct ldlm_reply *dlm_rep;
1126         struct ldlm_lock *lock;
1127         int rc;
1128         int size[2] = { [MSG_PTLRPC_BODY_OFF] = sizeof(struct ptlrpc_body),
1129                         [DLM_LOCKREPLY_OFF]   = sizeof(*dlm_rep) };
1130         ENTRY;
1131
1132         if (req->rq_export && req->rq_export->exp_ldlm_stats)
1133                 lprocfs_counter_incr(req->rq_export->exp_ldlm_stats,
1134                                      LDLM_CONVERT - LDLM_FIRST_OPC);
1135
1136         rc = lustre_pack_reply(req, 2, size, NULL);
1137         if (rc)
1138                 RETURN(rc);
1139
1140         dlm_rep = lustre_msg_buf(req->rq_repmsg, DLM_LOCKREPLY_OFF,
1141                                  sizeof(*dlm_rep));
1142         dlm_rep->lock_flags = dlm_req->lock_flags;
1143
1144         lock = ldlm_handle2lock(&dlm_req->lock_handle[0]);
1145         if (!lock) {
1146                 req->rq_status = EINVAL;
1147         } else {
1148                 void *res = NULL;
1149
1150                 LDLM_DEBUG(lock, "server-side convert handler START");
1151
1152                 do_gettimeofday(&lock->l_enqueued_time);
1153                 res = ldlm_lock_convert(lock, dlm_req->lock_desc.l_req_mode,
1154                                         &dlm_rep->lock_flags);
1155                 if (res) {
1156                         if (ldlm_del_waiting_lock(lock))
1157                                 LDLM_DEBUG(lock, "converted waiting lock");
1158                         req->rq_status = 0;
1159                 } else {
1160                         req->rq_status = EDEADLOCK;
1161                 }
1162         }
1163
1164         if (lock) {
1165                 if (!req->rq_status)
1166                         ldlm_reprocess_all(lock->l_resource);
1167                 LDLM_DEBUG(lock, "server-side convert handler END");
1168                 LDLM_LOCK_PUT(lock);
1169         } else
1170                 LDLM_DEBUG_NOLOCK("server-side convert handler END");
1171
1172         RETURN(0);
1173 }
1174
1175 int ldlm_handle_convert(struct ptlrpc_request *req)
1176 {
1177         int rc;
1178         struct ldlm_request *dlm_req;
1179
1180         dlm_req = lustre_swab_reqbuf(req, DLM_LOCKREQ_OFF, sizeof *dlm_req,
1181                                      lustre_swab_ldlm_request);
1182         if (dlm_req != NULL) {
1183                 rc = ldlm_handle_convert0(req, dlm_req);
1184         } else {
1185                 CERROR ("Can't unpack dlm_req\n");
1186                 rc = -EFAULT;
1187         }
1188         return rc;
1189 }
1190
1191 /* Cancel all the locks, which handles are packed into ldlm_request */
1192 int ldlm_request_cancel(struct ptlrpc_request *req,
1193                         const struct ldlm_request *dlm_req, int first)
1194 {
1195         struct ldlm_resource *res, *pres = NULL;
1196         struct ldlm_lock *lock;
1197         int i, count, done = 0;
1198         ENTRY;
1199
1200         count = dlm_req->lock_count ? dlm_req->lock_count : 1;
1201         if (first >= count)
1202                 RETURN(0);
1203
1204         /* There is no lock on the server at the replay time,
1205          * skip lock cancelling to make replay tests to pass. */
1206         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
1207                 RETURN(0);
1208
1209         LDLM_DEBUG_NOLOCK("server-side cancel handler START: %d locks, "
1210                           "starting at %d", count, first);
1211
1212         for (i = first; i < count; i++) {
1213                 lock = ldlm_handle2lock(&dlm_req->lock_handle[i]);
1214                 if (!lock) {
1215                         LDLM_DEBUG_NOLOCK("server-side cancel handler stale "
1216                                           "lock (cookie "LPU64")",
1217                                           dlm_req->lock_handle[i].cookie);
1218                         continue;
1219                 }
1220
1221                 res = lock->l_resource;
1222                 done++;
1223
1224                 if (res != pres) {
1225                         if (pres != NULL) {
1226                                 ldlm_reprocess_all(pres);
1227                                 ldlm_resource_putref(pres);
1228                         }
1229                         if (res != NULL) {
1230                                 ldlm_resource_getref(res);
1231                                 ldlm_res_lvbo_update(res, NULL, 0, 1);
1232                         }
1233                         pres = res;
1234                 }
1235                 ldlm_lock_cancel(lock);
1236                 LDLM_LOCK_PUT(lock);
1237         }
1238         if (pres != NULL) {
1239                 ldlm_reprocess_all(pres);
1240                 ldlm_resource_putref(pres);
1241         }
1242         LDLM_DEBUG_NOLOCK("server-side cancel handler END");
1243         RETURN(done);
1244 }
1245
1246 int ldlm_handle_cancel(struct ptlrpc_request *req)
1247 {
1248         struct ldlm_request *dlm_req;
1249         int rc;
1250         ENTRY;
1251
1252         dlm_req = lustre_swab_reqbuf(req, DLM_LOCKREQ_OFF, sizeof(*dlm_req),
1253                                      lustre_swab_ldlm_request);
1254         if (dlm_req == NULL) {
1255                 CERROR("bad request buffer for cancel\n");
1256                 RETURN(-EFAULT);
1257         }
1258
1259         if (req->rq_export && req->rq_export->exp_ldlm_stats)
1260                 lprocfs_counter_incr(req->rq_export->exp_ldlm_stats,
1261                                      LDLM_CANCEL - LDLM_FIRST_OPC);
1262
1263         rc = lustre_pack_reply(req, 1, NULL, NULL);
1264         if (rc)
1265                 RETURN(rc);
1266
1267         if (!ldlm_request_cancel(req, dlm_req, 0))
1268                 req->rq_status = ESTALE;
1269
1270         if (ptlrpc_reply(req) != 0)
1271                 LBUG();
1272
1273         RETURN(0);
1274 }
1275
1276 void ldlm_handle_bl_callback(struct ldlm_namespace *ns,
1277                              struct ldlm_lock_desc *ld, struct ldlm_lock *lock)
1278 {
1279         int do_ast;
1280         ENTRY;
1281
1282         LDLM_DEBUG(lock, "client blocking AST callback handler START");
1283
1284         lock_res_and_lock(lock);
1285         lock->l_flags |= LDLM_FL_CBPENDING;
1286
1287         if (lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK)
1288                 lock->l_flags |= LDLM_FL_CANCEL;
1289
1290         do_ast = (!lock->l_readers && !lock->l_writers);
1291         unlock_res_and_lock(lock);
1292
1293         if (do_ast) {
1294                 LDLM_DEBUG(lock, "already unused, calling "
1295                            "callback (%p)", lock->l_blocking_ast);
1296                 if (lock->l_blocking_ast != NULL)
1297                         lock->l_blocking_ast(lock, ld, lock->l_ast_data,
1298                                              LDLM_CB_BLOCKING);
1299         } else {
1300                 LDLM_DEBUG(lock, "Lock still has references, will be"
1301                            " cancelled later");
1302         }
1303
1304         LDLM_DEBUG(lock, "client blocking callback handler END");
1305         LDLM_LOCK_PUT(lock);
1306         EXIT;
1307 }
1308
1309 static void ldlm_handle_cp_callback(struct ptlrpc_request *req,
1310                                     struct ldlm_namespace *ns,
1311                                     struct ldlm_request *dlm_req,
1312                                     struct ldlm_lock *lock)
1313 {
1314         CFS_LIST_HEAD(ast_list);
1315         ENTRY;
1316
1317         LDLM_DEBUG(lock, "client completion callback handler START");
1318
1319         lock_res_and_lock(lock);
1320
1321         /* If we receive the completion AST before the actual enqueue returned,
1322          * then we might need to switch lock modes, resources, or extents. */
1323         if (dlm_req->lock_desc.l_granted_mode != lock->l_req_mode) {
1324                 lock->l_req_mode = dlm_req->lock_desc.l_granted_mode;
1325                 LDLM_DEBUG(lock, "completion AST, new lock mode");
1326         }
1327
1328         if (lock->l_resource->lr_type != LDLM_PLAIN) {
1329                 lock->l_policy_data = dlm_req->lock_desc.l_policy_data;
1330                 LDLM_DEBUG(lock, "completion AST, new policy data");
1331         }
1332
1333         ldlm_resource_unlink_lock(lock);
1334         if (memcmp(&dlm_req->lock_desc.l_resource.lr_name,
1335                    &lock->l_resource->lr_name,
1336                    sizeof(lock->l_resource->lr_name)) != 0) {
1337                 unlock_res_and_lock(lock);
1338                 ldlm_lock_change_resource(ns, lock,
1339                                         &dlm_req->lock_desc.l_resource.lr_name);
1340                 LDLM_DEBUG(lock, "completion AST, new resource");
1341                 CERROR("change resource!\n");
1342                 lock_res_and_lock(lock);
1343         }
1344
1345         if (dlm_req->lock_flags & LDLM_FL_AST_SENT) {
1346                 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_BL_AST;
1347                 LDLM_DEBUG(lock, "completion AST includes blocking AST");
1348         }
1349
1350         if (lock->l_lvb_len) {
1351                 void *lvb;
1352                 lvb = lustre_swab_reqbuf(req, DLM_REQ_REC_OFF, lock->l_lvb_len,
1353                                          lock->l_lvb_swabber);
1354                 if (lvb == NULL) {
1355                         LDLM_ERROR(lock, "completion AST did not contain "
1356                                    "expected LVB!");
1357                 } else {
1358                         memcpy(lock->l_lvb_data, lvb, lock->l_lvb_len);
1359                 }
1360         }
1361
1362         ldlm_grant_lock(lock, &ast_list);
1363         unlock_res_and_lock(lock);
1364
1365         LDLM_DEBUG(lock, "callback handler finished, about to run_ast_work");
1366
1367         ldlm_run_cp_ast_work(&ast_list);
1368
1369         LDLM_DEBUG_NOLOCK("client completion callback handler END (lock %p)",
1370                           lock);
1371         LDLM_LOCK_PUT(lock);
1372         EXIT;
1373 }
1374
1375 static void ldlm_handle_gl_callback(struct ptlrpc_request *req,
1376                                     struct ldlm_namespace *ns,
1377                                     struct ldlm_request *dlm_req,
1378                                     struct ldlm_lock *lock)
1379 {
1380         int rc = -ENOSYS;
1381         ENTRY;
1382
1383         LDLM_DEBUG(lock, "client glimpse AST callback handler");
1384
1385         if (lock->l_glimpse_ast != NULL)
1386                 rc = lock->l_glimpse_ast(lock, req);
1387
1388         if (req->rq_repmsg != NULL) {
1389                 ptlrpc_reply(req);
1390         } else {
1391                 req->rq_status = rc;
1392                 ptlrpc_error(req);
1393         }
1394
1395         lock_res_and_lock(lock);
1396         if (lock->l_granted_mode == LCK_PW &&
1397             !lock->l_readers && !lock->l_writers &&
1398             cfs_time_after(cfs_time_current(),
1399                            cfs_time_add(lock->l_last_used,
1400                                         cfs_time_seconds(10)))) {
1401                 unlock_res_and_lock(lock);
1402                 if (ldlm_bl_to_thread_lock(ns, NULL, lock))
1403                         ldlm_handle_bl_callback(ns, NULL, lock);
1404
1405                 EXIT;
1406                 return;
1407         }
1408         unlock_res_and_lock(lock);
1409         LDLM_LOCK_PUT(lock);
1410         EXIT;
1411 }
1412
1413 static int ldlm_callback_reply(struct ptlrpc_request *req, int rc)
1414 {
1415         req->rq_status = rc;
1416         if (req->rq_reply_state == NULL) {
1417                 rc = lustre_pack_reply(req, 1, NULL, NULL);
1418                 if (rc)
1419                         return rc;
1420         }
1421         return ptlrpc_reply(req);
1422 }
1423
1424 #ifdef __KERNEL__
1425 static int ldlm_bl_to_thread(struct ldlm_namespace *ns,
1426                              struct ldlm_lock_desc *ld, struct ldlm_lock *lock,
1427                              struct list_head *cancels, int count)
1428 {
1429         struct ldlm_bl_pool *blp = ldlm_state->ldlm_bl_pool;
1430         struct ldlm_bl_work_item *blwi;
1431         ENTRY;
1432
1433         if (cancels && count == 0)
1434                 RETURN(0);
1435
1436         OBD_ALLOC(blwi, sizeof(*blwi));
1437         if (blwi == NULL)
1438                 RETURN(-ENOMEM);
1439
1440         blwi->blwi_ns = ns;
1441         if (ld != NULL)
1442                 blwi->blwi_ld = *ld;
1443         if (count) {
1444                 list_add(&blwi->blwi_head, cancels);
1445                 list_del_init(cancels);
1446                 blwi->blwi_count = count;
1447         } else {
1448                 blwi->blwi_lock = lock;
1449         }
1450         spin_lock(&blp->blp_lock);
1451         if (lock && lock->l_flags & LDLM_FL_DISCARD_DATA) {
1452                 /* add LDLM_FL_DISCARD_DATA requests to the priority list */
1453                 list_add_tail(&blwi->blwi_entry, &blp->blp_prio_list);
1454         } else {
1455                 /* other blocking callbacks are added to the regular list */
1456                 list_add_tail(&blwi->blwi_entry, &blp->blp_list);
1457         }
1458         cfs_waitq_signal(&blp->blp_waitq);
1459         spin_unlock(&blp->blp_lock);
1460
1461         RETURN(0);
1462 }
1463 #endif
1464
1465 int ldlm_bl_to_thread_lock(struct ldlm_namespace *ns, struct ldlm_lock_desc *ld,
1466                            struct ldlm_lock *lock)
1467 {
1468 #ifdef __KERNEL__
1469         RETURN(ldlm_bl_to_thread(ns, ld, lock, NULL, 0));
1470 #else
1471         RETURN(-ENOSYS);
1472 #endif
1473 }
1474
1475 int ldlm_bl_to_thread_list(struct ldlm_namespace *ns, struct ldlm_lock_desc *ld,
1476                            struct list_head *cancels, int count)
1477 {
1478 #ifdef __KERNEL__
1479         RETURN(ldlm_bl_to_thread(ns, ld, NULL, cancels, count));
1480 #else
1481         RETURN(-ENOSYS);
1482 #endif
1483 }
1484
1485 static int ldlm_callback_handler(struct ptlrpc_request *req)
1486 {
1487         struct ldlm_namespace *ns;
1488         struct ldlm_request *dlm_req;
1489         struct ldlm_lock *lock;
1490         int rc;
1491         ENTRY;
1492
1493         /* Requests arrive in sender's byte order.  The ptlrpc service
1494          * handler has already checked and, if necessary, byte-swapped the
1495          * incoming request message body, but I am responsible for the
1496          * message buffers. */
1497
1498         if (req->rq_export == NULL) {
1499                 struct ldlm_request *dlm_req;
1500
1501                 CDEBUG(D_RPCTRACE, "operation %d from %s with bad "
1502                        "export cookie "LPX64"; this is "
1503                        "normal if this node rebooted with a lock held\n",
1504                        lustre_msg_get_opc(req->rq_reqmsg),
1505                        libcfs_id2str(req->rq_peer),
1506                        lustre_msg_get_handle(req->rq_reqmsg)->cookie);
1507
1508                 dlm_req = lustre_swab_reqbuf(req, DLM_LOCKREQ_OFF,
1509                                              sizeof(*dlm_req),
1510                                              lustre_swab_ldlm_request);
1511                 if (dlm_req != NULL)
1512                         CDEBUG(D_RPCTRACE, "--> lock cookie: "LPX64"\n",
1513                                dlm_req->lock_handle[0].cookie);
1514
1515                 ldlm_callback_reply(req, -ENOTCONN);
1516                 RETURN(0);
1517         }
1518
1519         LASSERT(req->rq_export != NULL);
1520         LASSERT(req->rq_export->exp_obd != NULL);
1521
1522         switch (lustre_msg_get_opc(req->rq_reqmsg)) {
1523         case LDLM_BL_CALLBACK:
1524                 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_BL_CALLBACK))
1525                         RETURN(0);
1526                 break;
1527         case LDLM_CP_CALLBACK:
1528                 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CP_CALLBACK))
1529                         RETURN(0);
1530                 break;
1531         case LDLM_GL_CALLBACK:
1532                 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_GL_CALLBACK))
1533                         RETURN(0);
1534                 break;
1535         case OBD_LOG_CANCEL: /* remove this eventually - for 1.4.0 compat */
1536                 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOG_CANCEL_NET))
1537                         RETURN(0);
1538                 rc = llog_origin_handle_cancel(req);
1539                 ldlm_callback_reply(req, rc);
1540                 RETURN(0);
1541         case OBD_QC_CALLBACK:
1542                 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_QC_CALLBACK_NET))
1543                         RETURN(0);
1544                 rc = target_handle_qc_callback(req);
1545                 ldlm_callback_reply(req, rc);
1546                 RETURN(0);
1547         case QUOTA_DQACQ:
1548         case QUOTA_DQREL:
1549                 /* reply in handler */
1550                 rc = target_handle_dqacq_callback(req);
1551                 RETURN(0);
1552         case LLOG_ORIGIN_HANDLE_CREATE:
1553                 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOGD_NET))
1554                         RETURN(0);
1555                 rc = llog_origin_handle_create(req);
1556                 ldlm_callback_reply(req, rc);
1557                 RETURN(0);
1558         case LLOG_ORIGIN_HANDLE_NEXT_BLOCK:
1559                 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOGD_NET))
1560                         RETURN(0);
1561                 rc = llog_origin_handle_next_block(req);
1562                 ldlm_callback_reply(req, rc);
1563                 RETURN(0);
1564         case LLOG_ORIGIN_HANDLE_READ_HEADER:
1565                 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOGD_NET))
1566                         RETURN(0);
1567                 rc = llog_origin_handle_read_header(req);
1568                 ldlm_callback_reply(req, rc);
1569                 RETURN(0);
1570         case LLOG_ORIGIN_HANDLE_CLOSE:
1571                 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOGD_NET))
1572                         RETURN(0);
1573                 rc = llog_origin_handle_close(req);
1574                 ldlm_callback_reply(req, rc);
1575                 RETURN(0);
1576         case SEC_CTX_FINI:
1577                 /* do nothing */
1578                 RETURN(0);
1579         default:
1580                 CERROR("unknown opcode %u\n",
1581                        lustre_msg_get_opc(req->rq_reqmsg));
1582                 ldlm_callback_reply(req, -EPROTO);
1583                 RETURN(0);
1584         }
1585
1586         ns = req->rq_export->exp_obd->obd_namespace;
1587         LASSERT(ns != NULL);
1588
1589         dlm_req = lustre_swab_reqbuf(req, DLM_LOCKREQ_OFF, sizeof(*dlm_req),
1590                                      lustre_swab_ldlm_request);
1591         if (dlm_req == NULL) {
1592                 CERROR ("can't unpack dlm_req\n");
1593                 ldlm_callback_reply(req, -EPROTO);
1594                 RETURN (0);
1595         }
1596
1597         lock = ldlm_handle2lock_ns(ns, &dlm_req->lock_handle[0]);
1598         if (!lock) {
1599                 CDEBUG(D_DLMTRACE, "callback on lock "LPX64" - lock "
1600                        "disappeared\n", dlm_req->lock_handle[0].cookie);
1601                 ldlm_callback_reply(req, -EINVAL);
1602                 RETURN(0);
1603         }
1604
1605         /* Copy hints/flags (e.g. LDLM_FL_DISCARD_DATA) from AST. */
1606         lock_res_and_lock(lock);
1607         lock->l_flags |= (dlm_req->lock_flags & LDLM_AST_FLAGS);
1608         if (lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK) {
1609                 /* If somebody cancels locks and cache is already droped,
1610                  * we can tell the server we have no lock. Otherwise, we
1611                  * should send cancel after dropping the cache. */
1612                 if ((lock->l_flags & LDLM_FL_CANCELING) &&
1613                     (lock->l_flags & LDLM_FL_BL_DONE)) {
1614                         LDLM_DEBUG(lock, "callback on lock "
1615                                    LPX64" - lock disappeared\n",
1616                                    dlm_req->lock_handle[0].cookie);
1617                         unlock_res_and_lock(lock);
1618                         LDLM_LOCK_PUT(lock);
1619                         ldlm_callback_reply(req, -EINVAL);
1620                         RETURN(0);
1621                 }
1622                 lock->l_flags |= LDLM_FL_BL_AST;
1623         }
1624         unlock_res_and_lock(lock);
1625
1626         /* We want the ost thread to get this reply so that it can respond
1627          * to ost requests (write cache writeback) that might be triggered
1628          * in the callback.
1629          *
1630          * But we'd also like to be able to indicate in the reply that we're
1631          * cancelling right now, because it's unused, or have an intent result
1632          * in the reply, so we might have to push the responsibility for sending
1633          * the reply down into the AST handlers, alas. */
1634
1635         switch (lustre_msg_get_opc(req->rq_reqmsg)) {
1636         case LDLM_BL_CALLBACK:
1637                 CDEBUG(D_INODE, "blocking ast\n");
1638                 if (!(lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK))
1639                         ldlm_callback_reply(req, 0);
1640                 if (ldlm_bl_to_thread_lock(ns, &dlm_req->lock_desc, lock))
1641                         ldlm_handle_bl_callback(ns, &dlm_req->lock_desc, lock);
1642                 break;
1643         case LDLM_CP_CALLBACK:
1644                 CDEBUG(D_INODE, "completion ast\n");
1645                 ldlm_callback_reply(req, 0);
1646                 ldlm_handle_cp_callback(req, ns, dlm_req, lock);
1647                 break;
1648         case LDLM_GL_CALLBACK:
1649                 CDEBUG(D_INODE, "glimpse ast\n");
1650                 ldlm_handle_gl_callback(req, ns, dlm_req, lock);
1651                 break;
1652         default:
1653                 LBUG();                         /* checked above */
1654         }
1655
1656         RETURN(0);
1657 }
1658
1659 static int ldlm_cancel_handler(struct ptlrpc_request *req)
1660 {
1661         int rc;
1662         ENTRY;
1663
1664         /* Requests arrive in sender's byte order.  The ptlrpc service
1665          * handler has already checked and, if necessary, byte-swapped the
1666          * incoming request message body, but I am responsible for the
1667          * message buffers. */
1668
1669         if (req->rq_export == NULL) {
1670                 struct ldlm_request *dlm_req;
1671
1672                 CERROR("operation %d from %s with bad export cookie "LPU64"\n",
1673                        lustre_msg_get_opc(req->rq_reqmsg),
1674                        libcfs_id2str(req->rq_peer),
1675                        lustre_msg_get_handle(req->rq_reqmsg)->cookie);
1676
1677                 dlm_req = lustre_swab_reqbuf(req, DLM_LOCKREQ_OFF,
1678                                              sizeof(*dlm_req),
1679                                              lustre_swab_ldlm_request);
1680                 if (dlm_req != NULL)
1681                         ldlm_lock_dump_handle(D_ERROR,
1682                                               &dlm_req->lock_handle[0]);
1683                 ldlm_callback_reply(req, -ENOTCONN);
1684                 RETURN(0);
1685         }
1686
1687         switch (lustre_msg_get_opc(req->rq_reqmsg)) {
1688
1689         /* XXX FIXME move this back to mds/handler.c, bug 249 */
1690         case LDLM_CANCEL:
1691                 CDEBUG(D_INODE, "cancel\n");
1692                 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL))
1693                         RETURN(0);
1694                 rc = ldlm_handle_cancel(req);
1695                 if (rc)
1696                         break;
1697                 RETURN(0);
1698         case OBD_LOG_CANCEL:
1699                 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOG_CANCEL_NET))
1700                         RETURN(0);
1701                 rc = llog_origin_handle_cancel(req);
1702                 ldlm_callback_reply(req, rc);
1703                 RETURN(0);
1704         default:
1705                 CERROR("invalid opcode %d\n",
1706                        lustre_msg_get_opc(req->rq_reqmsg));
1707                 ldlm_callback_reply(req, -EINVAL);
1708         }
1709
1710         RETURN(0);
1711 }
1712
1713 void ldlm_revoke_export_locks(struct obd_export *exp)
1714 {
1715         struct list_head *locklist = &exp->exp_ldlm_data.led_held_locks;
1716         struct list_head  rpc_list;
1717         struct ldlm_lock *lock, *next;
1718         struct ldlm_lock_desc desc;
1719
1720         ENTRY;
1721         INIT_LIST_HEAD(&rpc_list);
1722
1723         spin_lock(&exp->exp_ldlm_data.led_lock);
1724         list_for_each_entry_safe(lock, next, locklist, l_export_chain) {
1725                 lock_res_and_lock(lock);
1726
1727                 if (lock->l_req_mode != lock->l_granted_mode) {
1728                         unlock_res_and_lock(lock);
1729                         continue;
1730                 }
1731
1732                 LASSERT(lock->l_resource);
1733                 if (lock->l_resource->lr_type != LDLM_IBITS &&
1734                     lock->l_resource->lr_type != LDLM_PLAIN) {
1735                         unlock_res_and_lock(lock);
1736                         continue;
1737                 }
1738
1739                 if (lock->l_flags & LDLM_FL_AST_SENT) {
1740                         unlock_res_and_lock(lock);
1741                         continue;
1742                 }
1743
1744                 LASSERT(lock->l_blocking_ast);
1745                 LASSERT(!lock->l_blocking_lock);
1746
1747                 lock->l_flags |= LDLM_FL_AST_SENT;
1748                 list_move(&lock->l_export_chain, &rpc_list);
1749
1750                 unlock_res_and_lock(lock);
1751         }
1752         spin_unlock(&exp->exp_ldlm_data.led_lock);
1753
1754         while (!list_empty(&rpc_list)) {
1755                 lock = list_entry(rpc_list.next, struct ldlm_lock,
1756                                   l_export_chain);
1757                 list_del_init(&lock->l_export_chain);
1758
1759                 /* the desc just pretend to exclusive */
1760                 ldlm_lock2desc(lock, &desc);
1761                 desc.l_req_mode = LCK_EX;
1762                 desc.l_granted_mode = 0;
1763
1764                 LDLM_LOCK_GET(lock);
1765                 lock->l_blocking_ast(lock, &desc, lock->l_ast_data,
1766                                      LDLM_CB_BLOCKING);
1767                 LDLM_LOCK_PUT(lock);
1768         }
1769         EXIT;
1770 }
1771
1772 #ifdef __KERNEL__
1773 static struct ldlm_bl_work_item *ldlm_bl_get_work(struct ldlm_bl_pool *blp)
1774 {
1775         struct ldlm_bl_work_item *blwi = NULL;
1776         static unsigned int num_bl = 0;
1777
1778         spin_lock(&blp->blp_lock);
1779         /* process a request from the blp_list at least every blp_num_threads */
1780         if (!list_empty(&blp->blp_list) &&
1781             (list_empty(&blp->blp_prio_list) || num_bl == 0))
1782                 blwi = list_entry(blp->blp_list.next,
1783                                   struct ldlm_bl_work_item, blwi_entry);
1784         else
1785                 if (!list_empty(&blp->blp_prio_list))
1786                         blwi = list_entry(blp->blp_prio_list.next,
1787                                           struct ldlm_bl_work_item, blwi_entry);
1788
1789         if (blwi) {
1790                 if (++num_bl >= atomic_read(&blp->blp_num_threads))
1791                         num_bl = 0;
1792                 list_del(&blwi->blwi_entry);
1793         }
1794         spin_unlock(&blp->blp_lock);
1795
1796         return blwi;
1797 }
1798
1799 /* This only contains temporary data until the thread starts */
1800 struct ldlm_bl_thread_data {
1801         char                    bltd_name[CFS_CURPROC_COMM_MAX];
1802         struct ldlm_bl_pool     *bltd_blp;
1803         struct completion       bltd_comp;
1804         int                     bltd_num;
1805 };
1806
1807 static int ldlm_bl_thread_main(void *arg);
1808
1809 static int ldlm_bl_thread_start(struct ldlm_bl_pool *blp)
1810 {
1811         struct ldlm_bl_thread_data bltd = { .bltd_blp = blp };
1812         int rc;
1813
1814         init_completion(&bltd.bltd_comp);
1815         rc = cfs_kernel_thread(ldlm_bl_thread_main, &bltd, 0);
1816         if (rc < 0) {
1817                 CERROR("cannot start LDLM thread ldlm_bl_%02d: rc %d\n",
1818                        atomic_read(&blp->blp_num_threads), rc);
1819                 return rc;
1820         }
1821         wait_for_completion(&bltd.bltd_comp);
1822
1823         return 0;
1824 }
1825
1826 static int ldlm_bl_thread_main(void *arg)
1827 {
1828         struct ldlm_bl_pool *blp;
1829         ENTRY;
1830
1831         {
1832                 struct ldlm_bl_thread_data *bltd = arg;
1833
1834                 blp = bltd->bltd_blp;
1835
1836                 bltd->bltd_num = atomic_inc_return(&blp->blp_num_threads) - 1;
1837                 atomic_inc(&blp->blp_busy_threads);
1838
1839                 snprintf(bltd->bltd_name, sizeof(bltd->bltd_name) - 1,
1840                         "ldlm_bl_%02d", bltd->bltd_num);
1841                 cfs_daemonize(bltd->bltd_name);
1842
1843                 complete(&bltd->bltd_comp);
1844                 /* cannot use bltd after this, it is only on caller's stack */
1845         }
1846
1847         while (1) {
1848                 struct l_wait_info lwi = { 0 };
1849                 struct ldlm_bl_work_item *blwi = NULL;
1850
1851                 blwi = ldlm_bl_get_work(blp);
1852
1853                 if (blwi == NULL) {
1854                         int busy;
1855
1856                         atomic_dec(&blp->blp_busy_threads);
1857                         l_wait_event_exclusive(blp->blp_waitq,
1858                                          (blwi = ldlm_bl_get_work(blp)) != NULL,
1859                                          &lwi);
1860                         busy = atomic_inc_return(&blp->blp_busy_threads);
1861
1862                         if (blwi->blwi_ns == NULL)
1863                                 /* added by ldlm_cleanup() */
1864                                 break;
1865
1866                         /* Not fatal if racy and have a few too many threads */
1867                         if (unlikely(busy < blp->blp_max_threads &&
1868                                     busy >= atomic_read(&blp->blp_num_threads)))
1869                                 /* discard the return value, we tried */
1870                                 ldlm_bl_thread_start(blp);
1871                 } else {
1872                         if (blwi->blwi_ns == NULL)
1873                                 /* added by ldlm_cleanup() */
1874                                 break;
1875                 }
1876
1877                 if (blwi->blwi_count) {
1878                         /* The special case when we cancel locks in lru
1879                          * asynchronously, we pass the list of locks here.
1880                          * Thus lock is marked LDLM_FL_CANCELING, and already
1881                          * canceled locally. */
1882                         ldlm_cli_cancel_list(&blwi->blwi_head,
1883                                              blwi->blwi_count, NULL, 0, 0);
1884                 } else {
1885                         ldlm_handle_bl_callback(blwi->blwi_ns, &blwi->blwi_ld,
1886                                                 blwi->blwi_lock);
1887                 }
1888                 OBD_FREE(blwi, sizeof(*blwi));
1889         }
1890
1891         atomic_dec(&blp->blp_busy_threads);
1892         atomic_dec(&blp->blp_num_threads);
1893         complete(&blp->blp_comp);
1894         RETURN(0);
1895 }
1896
1897 #endif
1898
1899 static int ldlm_setup(void);
1900 static int ldlm_cleanup(int force);
1901
1902 int ldlm_get_ref(void)
1903 {
1904         int rc = 0;
1905         ENTRY;
1906         mutex_down(&ldlm_ref_sem);
1907         if (++ldlm_refcount == 1) {
1908                 rc = ldlm_setup();
1909                 if (rc)
1910                         ldlm_refcount--;
1911         }
1912         mutex_up(&ldlm_ref_sem);
1913
1914         RETURN(rc);
1915 }
1916
1917 void ldlm_put_ref(int force)
1918 {
1919         ENTRY;
1920         mutex_down(&ldlm_ref_sem);
1921         if (ldlm_refcount == 1) {
1922                 int rc = ldlm_cleanup(force);
1923                 if (rc)
1924                         CERROR("ldlm_cleanup failed: %d\n", rc);
1925                 else
1926                         ldlm_refcount--;
1927         } else {
1928                 ldlm_refcount--;
1929         }
1930         mutex_up(&ldlm_ref_sem);
1931
1932         EXIT;
1933 }
1934
1935 static int ldlm_setup(void)
1936 {
1937         struct ldlm_bl_pool *blp;
1938         int rc = 0;
1939         int ldlm_min_threads = LDLM_THREADS_AUTO_MIN;
1940         int ldlm_max_threads = LDLM_THREADS_AUTO_MAX;
1941 #ifdef __KERNEL__
1942         int i;
1943 #endif
1944         ENTRY;
1945
1946         if (ldlm_state != NULL)
1947                 RETURN(-EALREADY);
1948
1949         OBD_ALLOC(ldlm_state, sizeof(*ldlm_state));
1950         if (ldlm_state == NULL)
1951                 RETURN(-ENOMEM);
1952
1953 #ifdef LPROCFS
1954         rc = ldlm_proc_setup();
1955         if (rc != 0)
1956                 GOTO(out_free, rc);
1957 #endif
1958
1959 #ifdef __KERNEL__
1960         if (ldlm_num_threads) {
1961                 /* If ldlm_num_threads is set, it is the min and the max. */
1962                 if (ldlm_num_threads > LDLM_THREADS_AUTO_MAX)
1963                         ldlm_num_threads = LDLM_THREADS_AUTO_MAX;
1964                 if (ldlm_num_threads < LDLM_THREADS_AUTO_MIN)
1965                         ldlm_num_threads = LDLM_THREADS_AUTO_MIN;
1966                 ldlm_min_threads = ldlm_max_threads = ldlm_num_threads;
1967         }
1968 #endif
1969
1970         ldlm_state->ldlm_cb_service =
1971                 ptlrpc_init_svc(LDLM_NBUFS, LDLM_BUFSIZE, LDLM_MAXREQSIZE,
1972                                 LDLM_MAXREPSIZE, LDLM_CB_REQUEST_PORTAL,
1973                                 LDLM_CB_REPLY_PORTAL, ldlm_timeout * 900,
1974                                 ldlm_callback_handler, "ldlm_cbd",
1975                                 ldlm_svc_proc_dir, NULL,
1976                                 ldlm_min_threads, ldlm_max_threads,
1977                                 "ldlm_cb",
1978                                 LCT_MD_THREAD|LCT_DT_THREAD);
1979
1980         if (!ldlm_state->ldlm_cb_service) {
1981                 CERROR("failed to start service\n");
1982                 GOTO(out_proc, rc = -ENOMEM);
1983         }
1984
1985         ldlm_state->ldlm_cancel_service =
1986                 ptlrpc_init_svc(LDLM_NBUFS, LDLM_BUFSIZE, LDLM_MAXREQSIZE,
1987                                 LDLM_MAXREPSIZE, LDLM_CANCEL_REQUEST_PORTAL,
1988                                 LDLM_CANCEL_REPLY_PORTAL, ldlm_timeout * 6000,
1989                                 ldlm_cancel_handler, "ldlm_canceld",
1990                                 ldlm_svc_proc_dir, NULL,
1991                                 ldlm_min_threads, ldlm_max_threads,
1992                                 "ldlm_cn",
1993                                 LCT_MD_THREAD|LCT_DT_THREAD|LCT_CL_THREAD);
1994
1995         if (!ldlm_state->ldlm_cancel_service) {
1996                 CERROR("failed to start service\n");
1997                 GOTO(out_proc, rc = -ENOMEM);
1998         }
1999
2000         OBD_ALLOC(blp, sizeof(*blp));
2001         if (blp == NULL)
2002                 GOTO(out_proc, rc = -ENOMEM);
2003         ldlm_state->ldlm_bl_pool = blp;
2004
2005         spin_lock_init(&blp->blp_lock);
2006         CFS_INIT_LIST_HEAD(&blp->blp_list);
2007         CFS_INIT_LIST_HEAD(&blp->blp_prio_list);
2008         cfs_waitq_init(&blp->blp_waitq);
2009         atomic_set(&blp->blp_num_threads, 0);
2010         atomic_set(&blp->blp_busy_threads, 0);
2011         blp->blp_min_threads = ldlm_min_threads;
2012         blp->blp_max_threads = ldlm_max_threads;
2013
2014 #ifdef __KERNEL__
2015         for (i = 0; i < blp->blp_min_threads; i++) {
2016                 rc = ldlm_bl_thread_start(blp);
2017                 if (rc < 0)
2018                         GOTO(out_thread, rc);
2019         }
2020
2021         rc = ptlrpc_start_threads(NULL, ldlm_state->ldlm_cancel_service);
2022         if (rc)
2023                 GOTO(out_thread, rc);
2024
2025         rc = ptlrpc_start_threads(NULL, ldlm_state->ldlm_cb_service);
2026         if (rc)
2027                 GOTO(out_thread, rc);
2028
2029         CFS_INIT_LIST_HEAD(&expired_lock_thread.elt_expired_locks);
2030         expired_lock_thread.elt_state = ELT_STOPPED;
2031         cfs_waitq_init(&expired_lock_thread.elt_waitq);
2032
2033         CFS_INIT_LIST_HEAD(&waiting_locks_list);
2034         spin_lock_init(&waiting_locks_spinlock);
2035         cfs_timer_init(&waiting_locks_timer, waiting_locks_callback, 0);
2036
2037         rc = cfs_kernel_thread(expired_lock_main, NULL, CLONE_VM | CLONE_FILES);
2038         if (rc < 0) {
2039                 CERROR("Cannot start ldlm expired-lock thread: %d\n", rc);
2040                 GOTO(out_thread, rc);
2041         }
2042
2043         wait_event(expired_lock_thread.elt_waitq,
2044                    expired_lock_thread.elt_state == ELT_READY);
2045 #endif
2046
2047 #ifdef __KERNEL__
2048         rc = ldlm_pools_init();
2049         if (rc)
2050                 GOTO(out_thread, rc);
2051 #endif
2052         RETURN(0);
2053
2054 #ifdef __KERNEL__
2055  out_thread:
2056         ptlrpc_unregister_service(ldlm_state->ldlm_cancel_service);
2057         ptlrpc_unregister_service(ldlm_state->ldlm_cb_service);
2058 #endif
2059
2060  out_proc:
2061 #ifdef LPROCFS
2062         ldlm_proc_cleanup();
2063  out_free:
2064 #endif
2065         OBD_FREE(ldlm_state, sizeof(*ldlm_state));
2066         ldlm_state = NULL;
2067         return rc;
2068 }
2069
2070 static int ldlm_cleanup(int force)
2071 {
2072 #ifdef __KERNEL__
2073         struct ldlm_bl_pool *blp = ldlm_state->ldlm_bl_pool;
2074 #endif
2075         ENTRY;
2076
2077         if (!list_empty(ldlm_namespace_list(LDLM_NAMESPACE_SERVER)) || 
2078             !list_empty(ldlm_namespace_list(LDLM_NAMESPACE_CLIENT))) {
2079                 CERROR("ldlm still has namespaces; clean these up first.\n");
2080                 ldlm_dump_all_namespaces(LDLM_NAMESPACE_SERVER, D_DLMTRACE);
2081                 ldlm_dump_all_namespaces(LDLM_NAMESPACE_CLIENT, D_DLMTRACE);
2082                 RETURN(-EBUSY);
2083         }
2084
2085 #ifdef __KERNEL__
2086         ldlm_pools_fini();
2087 #endif
2088
2089 #ifdef __KERNEL__
2090         while (atomic_read(&blp->blp_num_threads) > 0) {
2091                 struct ldlm_bl_work_item blwi = { .blwi_ns = NULL };
2092
2093                 init_completion(&blp->blp_comp);
2094
2095                 spin_lock(&blp->blp_lock);
2096                 list_add_tail(&blwi.blwi_entry, &blp->blp_list);
2097                 cfs_waitq_signal(&blp->blp_waitq);
2098                 spin_unlock(&blp->blp_lock);
2099
2100                 wait_for_completion(&blp->blp_comp);
2101         }
2102         OBD_FREE(blp, sizeof(*blp));
2103
2104         ptlrpc_unregister_service(ldlm_state->ldlm_cb_service);
2105         ptlrpc_unregister_service(ldlm_state->ldlm_cancel_service);
2106         ldlm_proc_cleanup();
2107
2108         expired_lock_thread.elt_state = ELT_TERMINATE;
2109         cfs_waitq_signal(&expired_lock_thread.elt_waitq);
2110         wait_event(expired_lock_thread.elt_waitq,
2111                    expired_lock_thread.elt_state == ELT_STOPPED);
2112 #else
2113         ptlrpc_unregister_service(ldlm_state->ldlm_cb_service);
2114         ptlrpc_unregister_service(ldlm_state->ldlm_cancel_service);
2115 #endif
2116
2117         OBD_FREE(ldlm_state, sizeof(*ldlm_state));
2118         ldlm_state = NULL;
2119
2120         RETURN(0);
2121 }
2122
2123 int __init ldlm_init(void)
2124 {
2125         init_mutex(&ldlm_ref_sem);
2126         init_mutex(ldlm_namespace_lock(LDLM_NAMESPACE_SERVER));
2127         init_mutex(ldlm_namespace_lock(LDLM_NAMESPACE_CLIENT));
2128         ldlm_resource_slab = cfs_mem_cache_create("ldlm_resources",
2129                                                sizeof(struct ldlm_resource), 0,
2130                                                SLAB_HWCACHE_ALIGN);
2131         if (ldlm_resource_slab == NULL)
2132                 return -ENOMEM;
2133
2134         ldlm_lock_slab = cfs_mem_cache_create("ldlm_locks",
2135                                            sizeof(struct ldlm_lock), 0,
2136                                            SLAB_HWCACHE_ALIGN);
2137         if (ldlm_lock_slab == NULL) {
2138                 cfs_mem_cache_destroy(ldlm_resource_slab);
2139                 return -ENOMEM;
2140         }
2141
2142         return 0;
2143 }
2144
2145 void __exit ldlm_exit(void)
2146 {
2147         int rc;
2148         if (ldlm_refcount)
2149                 CERROR("ldlm_refcount is %d in ldlm_exit!\n", ldlm_refcount);
2150         rc = cfs_mem_cache_destroy(ldlm_resource_slab);
2151         LASSERTF(rc == 0, "couldn't free ldlm resource slab\n");
2152         rc = cfs_mem_cache_destroy(ldlm_lock_slab);
2153         LASSERTF(rc == 0, "couldn't free ldlm lock slab\n");
2154 }
2155
2156 /* ldlm_extent.c */
2157 EXPORT_SYMBOL(ldlm_extent_shift_kms);
2158
2159 /* ldlm_lock.c */
2160 EXPORT_SYMBOL(ldlm_get_processing_policy);
2161 EXPORT_SYMBOL(ldlm_lock2desc);
2162 EXPORT_SYMBOL(ldlm_register_intent);
2163 EXPORT_SYMBOL(ldlm_lockname);
2164 EXPORT_SYMBOL(ldlm_typename);
2165 EXPORT_SYMBOL(ldlm_lock2handle);
2166 EXPORT_SYMBOL(__ldlm_handle2lock);
2167 EXPORT_SYMBOL(ldlm_lock_get);
2168 EXPORT_SYMBOL(ldlm_lock_put);
2169 EXPORT_SYMBOL(ldlm_lock_match);
2170 EXPORT_SYMBOL(ldlm_lock_cancel);
2171 EXPORT_SYMBOL(ldlm_lock_addref);
2172 EXPORT_SYMBOL(ldlm_lock_decref);
2173 EXPORT_SYMBOL(ldlm_lock_decref_and_cancel);
2174 EXPORT_SYMBOL(ldlm_lock_change_resource);
2175 EXPORT_SYMBOL(ldlm_lock_set_data);
2176 EXPORT_SYMBOL(ldlm_it2str);
2177 EXPORT_SYMBOL(ldlm_lock_dump);
2178 EXPORT_SYMBOL(ldlm_lock_dump_handle);
2179 EXPORT_SYMBOL(ldlm_cancel_locks_for_export);
2180 EXPORT_SYMBOL(ldlm_reprocess_all_ns);
2181 EXPORT_SYMBOL(ldlm_lock_allow_match);
2182
2183 /* ldlm_request.c */
2184 EXPORT_SYMBOL(ldlm_completion_ast);
2185 EXPORT_SYMBOL(ldlm_blocking_ast);
2186 EXPORT_SYMBOL(ldlm_glimpse_ast);
2187 EXPORT_SYMBOL(ldlm_expired_completion_wait);
2188 EXPORT_SYMBOL(ldlm_prep_enqueue_req);
2189 EXPORT_SYMBOL(ldlm_prep_elc_req);
2190 EXPORT_SYMBOL(ldlm_cli_convert);
2191 EXPORT_SYMBOL(ldlm_cli_enqueue);
2192 EXPORT_SYMBOL(ldlm_cli_enqueue_fini);
2193 EXPORT_SYMBOL(ldlm_cli_enqueue_local);
2194 EXPORT_SYMBOL(ldlm_cli_cancel);
2195 EXPORT_SYMBOL(ldlm_cli_cancel_unused);
2196 EXPORT_SYMBOL(ldlm_cli_cancel_unused_resource);
2197 EXPORT_SYMBOL(ldlm_cli_cancel_req);
2198 EXPORT_SYMBOL(ldlm_cli_join_lru);
2199 EXPORT_SYMBOL(ldlm_replay_locks);
2200 EXPORT_SYMBOL(ldlm_resource_foreach);
2201 EXPORT_SYMBOL(ldlm_namespace_foreach);
2202 EXPORT_SYMBOL(ldlm_namespace_foreach_res);
2203 EXPORT_SYMBOL(ldlm_resource_iterate);
2204 EXPORT_SYMBOL(ldlm_cancel_resource_local);
2205 EXPORT_SYMBOL(ldlm_cli_cancel_list);
2206
2207 /* ldlm_lockd.c */
2208 EXPORT_SYMBOL(ldlm_server_blocking_ast);
2209 EXPORT_SYMBOL(ldlm_server_completion_ast);
2210 EXPORT_SYMBOL(ldlm_server_glimpse_ast);
2211 EXPORT_SYMBOL(ldlm_handle_enqueue);
2212 EXPORT_SYMBOL(ldlm_handle_enqueue0);
2213 EXPORT_SYMBOL(ldlm_handle_cancel);
2214 EXPORT_SYMBOL(ldlm_request_cancel);
2215 EXPORT_SYMBOL(ldlm_handle_convert);
2216 EXPORT_SYMBOL(ldlm_handle_convert0);
2217 EXPORT_SYMBOL(ldlm_del_waiting_lock);
2218 EXPORT_SYMBOL(ldlm_get_ref);
2219 EXPORT_SYMBOL(ldlm_put_ref);
2220 EXPORT_SYMBOL(ldlm_refresh_waiting_lock);
2221 EXPORT_SYMBOL(ldlm_revoke_export_locks);
2222
2223 /* ldlm_resource.c */
2224 EXPORT_SYMBOL(ldlm_namespace_new);
2225 EXPORT_SYMBOL(ldlm_namespace_cleanup);
2226 EXPORT_SYMBOL(ldlm_namespace_free);
2227 EXPORT_SYMBOL(ldlm_namespace_dump);
2228 EXPORT_SYMBOL(ldlm_dump_all_namespaces);
2229 EXPORT_SYMBOL(ldlm_resource_get);
2230 EXPORT_SYMBOL(ldlm_resource_putref);
2231 EXPORT_SYMBOL(ldlm_resource_unlink_lock);
2232
2233 /* ldlm_lib.c */
2234 EXPORT_SYMBOL(client_import_add_conn);
2235 EXPORT_SYMBOL(client_import_del_conn);
2236 EXPORT_SYMBOL(client_obd_setup);
2237 EXPORT_SYMBOL(client_obd_cleanup);
2238 EXPORT_SYMBOL(client_connect_import);
2239 EXPORT_SYMBOL(client_disconnect_export);
2240 EXPORT_SYMBOL(target_start_recovery_thread);
2241 EXPORT_SYMBOL(target_stop_recovery_thread);
2242 EXPORT_SYMBOL(target_handle_connect);
2243 EXPORT_SYMBOL(target_cleanup_recovery);
2244 EXPORT_SYMBOL(target_destroy_export);
2245 EXPORT_SYMBOL(target_cancel_recovery_timer);
2246 EXPORT_SYMBOL(target_send_reply);
2247 EXPORT_SYMBOL(target_queue_recovery_request);
2248 EXPORT_SYMBOL(target_handle_ping);
2249 EXPORT_SYMBOL(target_pack_pool_reply);
2250 EXPORT_SYMBOL(target_handle_disconnect);
2251
2252 /* l_lock.c */
2253 EXPORT_SYMBOL(lock_res_and_lock);
2254 EXPORT_SYMBOL(unlock_res_and_lock);
2255