Whamcloud - gitweb
e140e0fdcf71fce198d1c115d394d8c52e12ed27
[fs/lustre-release.git] / lustre / ldlm / ldlm_request.c
1 /*
2  * GPL HEADER START
3  *
4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
5  *
6  * This program is free software; you can redistribute it and/or modify
7  * it under the terms of the GNU General Public License version 2 only,
8  * as published by the Free Software Foundation.
9  *
10  * This program is distributed in the hope that it will be useful, but
11  * WITHOUT ANY WARRANTY; without even the implied warranty of
12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
13  * General Public License version 2 for more details (a copy is included
14  * in the LICENSE file that accompanied this code).
15  *
16  * You should have received a copy of the GNU General Public License
17  * version 2 along with this program; If not, see
18  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
19  *
20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21  * CA 95054 USA or visit www.sun.com if you need additional information or
22  * have any questions.
23  *
24  * GPL HEADER END
25  */
26 /*
27  * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
28  * Use is subject to license terms.
29  *
30  * Copyright (c) 2010, 2013, Intel Corporation.
31  */
32 /*
33  * This file is part of Lustre, http://www.lustre.org/
34  * Lustre is a trademark of Sun Microsystems, Inc.
35  */
36 /**
37  * This file contains Asynchronous System Trap (AST) handlers and related
38  * LDLM request-processing routines.
39  *
40  * An AST is a callback issued on a lock when its state is changed. There are
41  * several different types of ASTs (callbacks) registered for each lock:
42  *
43  * - completion AST: when a lock is enqueued by some process, but cannot be
44  *   granted immediately due to other conflicting locks on the same resource,
45  *   the completion AST is sent to notify the caller when the lock is
46  *   eventually granted
47  *
48  * - blocking AST: when a lock is granted to some process, if another process
49  *   enqueues a conflicting (blocking) lock on a resource, a blocking AST is
50  *   sent to notify the holder(s) of the lock(s) of the conflicting lock
51  *   request. The lock holder(s) must release their lock(s) on that resource in
52  *   a timely manner or be evicted by the server.
53  *
54  * - glimpse AST: this is used when a process wants information about a lock
55  *   (i.e. the lock value block (LVB)) but does not necessarily require holding
56  *   the lock. If the resource is locked, the lock holder(s) are sent glimpse
57  *   ASTs and the LVB is returned to the caller, and lock holder(s) may CANCEL
58  *   their lock(s) if they are idle. If the resource is not locked, the server
59  *   may grant the lock.
60  */
61
62 #define DEBUG_SUBSYSTEM S_LDLM
63 #ifndef __KERNEL__
64 #include <signal.h>
65 #include <liblustre.h>
66 #endif
67
68 #include <lustre_dlm.h>
69 #include <obd_class.h>
70 #include <obd.h>
71
72 #include "ldlm_internal.h"
73
74 int ldlm_enqueue_min = OBD_TIMEOUT_DEFAULT;
75 CFS_MODULE_PARM(ldlm_enqueue_min, "i", int, 0644,
76                 "lock enqueue timeout minimum");
77
78 /* in client side, whether the cached locks will be canceled before replay */
79 unsigned int ldlm_cancel_unused_locks_before_replay = 1;
80
81 static void interrupted_completion_wait(void *data)
82 {
83 }
84
85 struct lock_wait_data {
86         struct ldlm_lock *lwd_lock;
87         __u32             lwd_conn_cnt;
88 };
89
90 struct ldlm_async_args {
91         struct lustre_handle lock_handle;
92 };
93
94 int ldlm_expired_completion_wait(void *data)
95 {
96         struct lock_wait_data *lwd = data;
97         struct ldlm_lock *lock = lwd->lwd_lock;
98         struct obd_import *imp;
99         struct obd_device *obd;
100
101         ENTRY;
102         if (lock->l_conn_export == NULL) {
103                 static cfs_time_t next_dump = 0, last_dump = 0;
104
105                 LCONSOLE_WARN("lock timed out (enqueued at "CFS_TIME_T", "
106                               CFS_DURATION_T"s ago)\n",
107                               lock->l_last_activity,
108                               cfs_time_sub(cfs_time_current_sec(),
109                                            lock->l_last_activity));
110                 LDLM_DEBUG(lock, "lock timed out (enqueued at "CFS_TIME_T", "
111                            CFS_DURATION_T"s ago); not entering recovery in "
112                            "server code, just going back to sleep",
113                            lock->l_last_activity,
114                            cfs_time_sub(cfs_time_current_sec(),
115                                         lock->l_last_activity));
116                 if (cfs_time_after(cfs_time_current(), next_dump)) {
117                         last_dump = next_dump;
118                         next_dump = cfs_time_shift(300);
119                         ldlm_namespace_dump(D_DLMTRACE,
120                                             ldlm_lock_to_ns(lock));
121                         if (last_dump == 0)
122                                 libcfs_debug_dumplog();
123                 }
124                 RETURN(0);
125         }
126
127         obd = lock->l_conn_export->exp_obd;
128         imp = obd->u.cli.cl_import;
129         ptlrpc_fail_import(imp, lwd->lwd_conn_cnt);
130         LDLM_ERROR(lock, "lock timed out (enqueued at "CFS_TIME_T", "
131                   CFS_DURATION_T"s ago), entering recovery for %s@%s",
132                   lock->l_last_activity,
133                   cfs_time_sub(cfs_time_current_sec(), lock->l_last_activity),
134                   obd2cli_tgt(obd), imp->imp_connection->c_remote_uuid.uuid);
135
136         RETURN(0);
137 }
138 EXPORT_SYMBOL(ldlm_expired_completion_wait);
139
140 /* We use the same basis for both server side and client side functions
141    from a single node. */
142 int ldlm_get_enq_timeout(struct ldlm_lock *lock)
143 {
144         int timeout = at_get(ldlm_lock_to_ns_at(lock));
145         if (AT_OFF)
146                 return obd_timeout / 2;
147         /* Since these are non-updating timeouts, we should be conservative.
148            It would be nice to have some kind of "early reply" mechanism for
149            lock callbacks too... */
150         timeout = min_t(int, at_max, timeout + (timeout >> 1)); /* 150% */
151         return max(timeout, ldlm_enqueue_min);
152 }
153 EXPORT_SYMBOL(ldlm_get_enq_timeout);
154
155 /**
156  * Helper function for ldlm_completion_ast(), updating timings when lock is
157  * actually granted.
158  */
159 static int ldlm_completion_tail(struct ldlm_lock *lock)
160 {
161         long delay;
162         int  result;
163
164         if (ldlm_is_destroyed(lock) || ldlm_is_failed(lock)) {
165                 LDLM_DEBUG(lock, "client-side enqueue: destroyed");
166                 result = -EIO;
167         } else {
168                 delay = cfs_time_sub(cfs_time_current_sec(),
169                                      lock->l_last_activity);
170                 LDLM_DEBUG(lock, "client-side enqueue: granted after "
171                            CFS_DURATION_T"s", delay);
172
173                 /* Update our time estimate */
174                 at_measured(ldlm_lock_to_ns_at(lock),
175                             delay);
176                 result = 0;
177         }
178         return result;
179 }
180
181 /**
182  * Implementation of ->l_completion_ast() for a client, that doesn't wait
183  * until lock is granted. Suitable for locks enqueued through ptlrpcd, of
184  * other threads that cannot block for long.
185  */
186 int ldlm_completion_ast_async(struct ldlm_lock *lock, __u64 flags, void *data)
187 {
188         ENTRY;
189
190         if (flags == LDLM_FL_WAIT_NOREPROC) {
191                 LDLM_DEBUG(lock, "client-side enqueue waiting on pending lock");
192                 RETURN(0);
193         }
194
195         if (!(flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED |
196                        LDLM_FL_BLOCK_CONV))) {
197                 wake_up(&lock->l_waitq);
198                 RETURN(ldlm_completion_tail(lock));
199         }
200
201         LDLM_DEBUG(lock, "client-side enqueue returned a blocked lock, "
202                    "going forward");
203         ldlm_reprocess_all(lock->l_resource);
204         RETURN(0);
205 }
206 EXPORT_SYMBOL(ldlm_completion_ast_async);
207
208 /**
209  * Generic LDLM "completion" AST. This is called in several cases:
210  *
211  *     - when a reply to an ENQUEUE RPC is received from the server
212  *       (ldlm_cli_enqueue_fini()). Lock might be granted or not granted at
213  *       this point (determined by flags);
214  *
215  *     - when LDLM_CP_CALLBACK RPC comes to client to notify it that lock has
216  *       been granted;
217  *
218  *     - when ldlm_lock_match(LDLM_FL_LVB_READY) is about to wait until lock
219  *       gets correct lvb;
220  *
221  *     - to force all locks when resource is destroyed (cleanup_resource());
222  *
223  *     - during lock conversion (not used currently).
224  *
225  * If lock is not granted in the first case, this function waits until second
226  * or penultimate cases happen in some other thread.
227  *
228  */
229 int ldlm_completion_ast(struct ldlm_lock *lock, __u64 flags, void *data)
230 {
231         /* XXX ALLOCATE - 160 bytes */
232         struct lock_wait_data lwd;
233         struct obd_device *obd;
234         struct obd_import *imp = NULL;
235         struct l_wait_info lwi;
236         __u32 timeout;
237         int rc = 0;
238         ENTRY;
239
240         if (flags == LDLM_FL_WAIT_NOREPROC) {
241                 LDLM_DEBUG(lock, "client-side enqueue waiting on pending lock");
242                 goto noreproc;
243         }
244
245         if (!(flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED |
246                        LDLM_FL_BLOCK_CONV))) {
247                 wake_up(&lock->l_waitq);
248                 RETURN(0);
249         }
250
251         LDLM_DEBUG(lock, "client-side enqueue returned a blocked lock, "
252                    "sleeping");
253
254 noreproc:
255
256         obd = class_exp2obd(lock->l_conn_export);
257
258         /* if this is a local lock, then there is no import */
259         if (obd != NULL) {
260                 imp = obd->u.cli.cl_import;
261         }
262
263         /* Wait a long time for enqueue - server may have to callback a
264            lock from another client.  Server will evict the other client if it
265            doesn't respond reasonably, and then give us the lock. */
266         timeout = ldlm_get_enq_timeout(lock) * 2;
267
268         lwd.lwd_lock = lock;
269
270         if (ldlm_is_no_timeout(lock)) {
271                 LDLM_DEBUG(lock, "waiting indefinitely because of NO_TIMEOUT");
272                 lwi = LWI_INTR(interrupted_completion_wait, &lwd);
273         } else {
274                 lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(timeout),
275                                        ldlm_expired_completion_wait,
276                                        interrupted_completion_wait, &lwd);
277         }
278
279         if (imp != NULL) {
280                 spin_lock(&imp->imp_lock);
281                 lwd.lwd_conn_cnt = imp->imp_conn_cnt;
282                 spin_unlock(&imp->imp_lock);
283         }
284
285         if (ns_is_client(ldlm_lock_to_ns(lock)) &&
286             OBD_FAIL_CHECK_RESET(OBD_FAIL_LDLM_INTR_CP_AST,
287                                  OBD_FAIL_LDLM_CP_BL_RACE | OBD_FAIL_ONCE)) {
288                 ldlm_set_fail_loc(lock);
289                 rc = -EINTR;
290         } else {
291                 /* Go to sleep until the lock is granted or cancelled. */
292                 rc = l_wait_event(lock->l_waitq,
293                                   is_granted_or_cancelled(lock), &lwi);
294         }
295
296         if (rc) {
297                 LDLM_DEBUG(lock, "client-side enqueue waking up: failed (%d)",
298                            rc);
299                 RETURN(rc);
300         }
301
302         RETURN(ldlm_completion_tail(lock));
303 }
304 EXPORT_SYMBOL(ldlm_completion_ast);
305
306 /**
307  * A helper to build a blocking AST function
308  *
309  * Perform a common operation for blocking ASTs:
310  * defferred lock cancellation.
311  *
312  * \param lock the lock blocking or canceling AST was called on
313  * \retval 0
314  * \see mdt_blocking_ast
315  * \see ldlm_blocking_ast
316  */
317 int ldlm_blocking_ast_nocheck(struct ldlm_lock *lock)
318 {
319         int do_ast;
320         ENTRY;
321
322         ldlm_set_cbpending(lock);
323         do_ast = (!lock->l_readers && !lock->l_writers);
324         unlock_res_and_lock(lock);
325
326         if (do_ast) {
327                 struct lustre_handle lockh;
328                 int rc;
329
330                 LDLM_DEBUG(lock, "already unused, calling ldlm_cli_cancel");
331                 ldlm_lock2handle(lock, &lockh);
332                 rc = ldlm_cli_cancel(&lockh, LCF_ASYNC);
333                 if (rc < 0)
334                         CERROR("ldlm_cli_cancel: %d\n", rc);
335         } else {
336                 LDLM_DEBUG(lock, "Lock still has references, will be "
337                            "cancelled later");
338         }
339         RETURN(0);
340 }
341 EXPORT_SYMBOL(ldlm_blocking_ast_nocheck);
342
343 /**
344  * Server blocking AST
345  *
346  * ->l_blocking_ast() callback for LDLM locks acquired by server-side
347  * OBDs.
348  *
349  * \param lock the lock which blocks a request or cancelling lock
350  * \param desc unused
351  * \param data unused
352  * \param flag indicates whether this cancelling or blocking callback
353  * \retval 0
354  * \see ldlm_blocking_ast_nocheck
355  */
356 int ldlm_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
357                       void *data, int flag)
358 {
359         ENTRY;
360
361         if (flag == LDLM_CB_CANCELING) {
362                 /* Don't need to do anything here. */
363                 RETURN(0);
364         }
365
366         lock_res_and_lock(lock);
367         /* Get this: if ldlm_blocking_ast is racing with intent_policy, such
368          * that ldlm_blocking_ast is called just before intent_policy method
369          * takes the lr_lock, then by the time we get the lock, we might not
370          * be the correct blocking function anymore.  So check, and return
371          * early, if so. */
372         if (lock->l_blocking_ast != ldlm_blocking_ast) {
373                 unlock_res_and_lock(lock);
374                 RETURN(0);
375         }
376         RETURN(ldlm_blocking_ast_nocheck(lock));
377 }
378 EXPORT_SYMBOL(ldlm_blocking_ast);
379
380 /**
381  * ->l_glimpse_ast() for DLM extent locks acquired on the server-side. See
382  * comment in filter_intent_policy() on why you may need this.
383  */
384 int ldlm_glimpse_ast(struct ldlm_lock *lock, void *reqp)
385 {
386         /*
387          * Returning -ELDLM_NO_LOCK_DATA actually works, but the reason for
388          * that is rather subtle: with OST-side locking, it may so happen that
389          * _all_ extent locks are held by the OST. If client wants to obtain
390          * current file size it calls ll{,u}_glimpse_size(), and (as locks are
391          * on the server), dummy glimpse callback fires and does
392          * nothing. Client still receives correct file size due to the
393          * following fragment in filter_intent_policy():
394          *
395          * rc = l->l_glimpse_ast(l, NULL); // this will update the LVB
396          * if (rc != 0 && res->lr_namespace->ns_lvbo &&
397          *     res->lr_namespace->ns_lvbo->lvbo_update) {
398          *         res->lr_namespace->ns_lvbo->lvbo_update(res, NULL, 0, 1);
399          * }
400          *
401          * that is, after glimpse_ast() fails, filter_lvbo_update() runs, and
402          * returns correct file size to the client.
403          */
404         return -ELDLM_NO_LOCK_DATA;
405 }
406 EXPORT_SYMBOL(ldlm_glimpse_ast);
407
408 /**
409  * Enqueue a local lock (typically on a server).
410  */
411 int ldlm_cli_enqueue_local(struct ldlm_namespace *ns,
412                            const struct ldlm_res_id *res_id,
413                            ldlm_type_t type, ldlm_policy_data_t *policy,
414                            ldlm_mode_t mode, __u64 *flags,
415                            ldlm_blocking_callback blocking,
416                            ldlm_completion_callback completion,
417                            ldlm_glimpse_callback glimpse,
418                            void *data, __u32 lvb_len, enum lvb_type lvb_type,
419                            const __u64 *client_cookie,
420                            struct lustre_handle *lockh)
421 {
422         struct ldlm_lock *lock;
423         int err;
424         const struct ldlm_callback_suite cbs = { .lcs_completion = completion,
425                                                  .lcs_blocking   = blocking,
426                                                  .lcs_glimpse    = glimpse,
427         };
428         ENTRY;
429
430         LASSERT(!(*flags & LDLM_FL_REPLAY));
431         if (unlikely(ns_is_client(ns))) {
432                 CERROR("Trying to enqueue local lock in a shadow namespace\n");
433                 LBUG();
434         }
435
436         lock = ldlm_lock_create(ns, res_id, type, mode, &cbs, data, lvb_len,
437                                 lvb_type);
438         if (unlikely(!lock))
439                 GOTO(out_nolock, err = -ENOMEM);
440
441         ldlm_lock2handle(lock, lockh);
442
443         /* NB: we don't have any lock now (lock_res_and_lock)
444          * because it's a new lock */
445         ldlm_lock_addref_internal_nolock(lock, mode);
446         ldlm_set_local(lock);
447         if (*flags & LDLM_FL_ATOMIC_CB)
448                 ldlm_set_atomic_cb(lock);
449
450         if (policy != NULL)
451                 lock->l_policy_data = *policy;
452         if (client_cookie != NULL)
453                 lock->l_client_cookie = *client_cookie;
454         if (type == LDLM_EXTENT) {
455                 /* extent lock without policy is a bug */
456                 if (policy == NULL)
457                         LBUG();
458
459                 lock->l_req_extent = policy->l_extent;
460         }
461
462         err = ldlm_lock_enqueue(ns, &lock, policy, flags);
463         if (unlikely(err != ELDLM_OK))
464                 GOTO(out, err);
465
466         if (policy != NULL)
467                 *policy = lock->l_policy_data;
468
469         if (lock->l_completion_ast)
470                 lock->l_completion_ast(lock, *flags, NULL);
471
472         LDLM_DEBUG(lock, "client-side local enqueue handler, new lock created");
473         EXIT;
474  out:
475         LDLM_LOCK_RELEASE(lock);
476  out_nolock:
477         return err;
478 }
479 EXPORT_SYMBOL(ldlm_cli_enqueue_local);
480
481 static void failed_lock_cleanup(struct ldlm_namespace *ns,
482                                 struct ldlm_lock *lock, int mode)
483 {
484         int need_cancel = 0;
485
486         /* Set a flag to prevent us from sending a CANCEL (bug 407) */
487         lock_res_and_lock(lock);
488         /* Check that lock is not granted or failed, we might race. */
489         if ((lock->l_req_mode != lock->l_granted_mode) &&
490             !ldlm_is_failed(lock)) {
491                 /* Make sure that this lock will not be found by raced
492                  * bl_ast and -EINVAL reply is sent to server anyways.
493                  * b=17645*/
494                 lock->l_flags |= LDLM_FL_LOCAL_ONLY | LDLM_FL_FAILED |
495                                  LDLM_FL_ATOMIC_CB | LDLM_FL_CBPENDING;
496                 need_cancel = 1;
497         }
498         unlock_res_and_lock(lock);
499
500         if (need_cancel)
501                 LDLM_DEBUG(lock,
502                            "setting FL_LOCAL_ONLY | LDLM_FL_FAILED | "
503                            "LDLM_FL_ATOMIC_CB | LDLM_FL_CBPENDING");
504         else
505                 LDLM_DEBUG(lock, "lock was granted or failed in race");
506
507         ldlm_lock_decref_internal(lock, mode);
508
509         /* XXX - HACK because we shouldn't call ldlm_lock_destroy()
510          *       from llite/file.c/ll_file_flock(). */
511         /* This code makes for the fact that we do not have blocking handler on
512          * a client for flock locks. As such this is the place where we must
513          * completely kill failed locks. (interrupted and those that
514          * were waiting to be granted when server evicted us. */
515         if (lock->l_resource->lr_type == LDLM_FLOCK) {
516                 lock_res_and_lock(lock);
517                 ldlm_resource_unlink_lock(lock);
518                 ldlm_lock_destroy_nolock(lock);
519                 unlock_res_and_lock(lock);
520         }
521 }
522
523 /**
524  * Finishing portion of client lock enqueue code.
525  *
526  * Called after receiving reply from server.
527  */
528 int ldlm_cli_enqueue_fini(struct obd_export *exp, struct ptlrpc_request *req,
529                           ldlm_type_t type, __u8 with_policy, ldlm_mode_t mode,
530                           __u64 *flags, void *lvb, __u32 lvb_len,
531                           struct lustre_handle *lockh,int rc)
532 {
533         struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
534         int is_replay = *flags & LDLM_FL_REPLAY;
535         struct ldlm_lock *lock;
536         struct ldlm_reply *reply;
537         int cleanup_phase = 1;
538         int size = 0;
539         ENTRY;
540
541         lock = ldlm_handle2lock(lockh);
542         /* ldlm_cli_enqueue is holding a reference on this lock. */
543         if (!lock) {
544                 LASSERT(type == LDLM_FLOCK);
545                 RETURN(-ENOLCK);
546         }
547
548         LASSERTF(ergo(lvb_len != 0, lvb_len == lock->l_lvb_len),
549                  "lvb_len = %d, l_lvb_len = %d\n", lvb_len, lock->l_lvb_len);
550
551         if (rc != ELDLM_OK) {
552                 LASSERT(!is_replay);
553                 LDLM_DEBUG(lock, "client-side enqueue END (%s)",
554                            rc == ELDLM_LOCK_ABORTED ? "ABORTED" : "FAILED");
555
556                 if (rc != ELDLM_LOCK_ABORTED)
557                         GOTO(cleanup, rc);
558         }
559
560         /* Before we return, swab the reply */
561         reply = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
562         if (reply == NULL)
563                 GOTO(cleanup, rc = -EPROTO);
564
565         if (lvb_len != 0) {
566                 LASSERT(lvb != NULL);
567
568                 size = req_capsule_get_size(&req->rq_pill, &RMF_DLM_LVB,
569                                             RCL_SERVER);
570                 if (size < 0) {
571                         LDLM_ERROR(lock, "Fail to get lvb_len, rc = %d", size);
572                         GOTO(cleanup, rc = size);
573                 } else if (unlikely(size > lvb_len)) {
574                         LDLM_ERROR(lock, "Replied LVB is larger than "
575                                    "expectation, expected = %d, replied = %d",
576                                    lvb_len, size);
577                         GOTO(cleanup, rc = -EINVAL);
578                 }
579         }
580
581         if (rc == ELDLM_LOCK_ABORTED) {
582                 if (lvb_len != 0)
583                         rc = ldlm_fill_lvb(lock, &req->rq_pill, RCL_SERVER,
584                                            lvb, size);
585                 GOTO(cleanup, rc = (rc != 0 ? rc : ELDLM_LOCK_ABORTED));
586         }
587
588         /* lock enqueued on the server */
589         cleanup_phase = 0;
590
591         lock_res_and_lock(lock);
592         /* Key change rehash lock in per-export hash with new key */
593         if (exp->exp_lock_hash) {
594                 /* In the function below, .hs_keycmp resolves to
595                  * ldlm_export_lock_keycmp() */
596                 /* coverity[overrun-buffer-val] */
597                 cfs_hash_rehash_key(exp->exp_lock_hash,
598                                     &lock->l_remote_handle,
599                                     &reply->lock_handle,
600                                     &lock->l_exp_hash);
601         } else {
602                 lock->l_remote_handle = reply->lock_handle;
603         }
604
605         *flags = ldlm_flags_from_wire(reply->lock_flags);
606         lock->l_flags |= ldlm_flags_from_wire(reply->lock_flags &
607                                               LDLM_FL_INHERIT_MASK);
608         /* move NO_TIMEOUT flag to the lock to force ldlm_lock_match()
609          * to wait with no timeout as well */
610         lock->l_flags |= ldlm_flags_from_wire(reply->lock_flags &
611                                               LDLM_FL_NO_TIMEOUT);
612         unlock_res_and_lock(lock);
613
614         CDEBUG(D_INFO, "local: %p, remote cookie: "LPX64", flags: 0x%llx\n",
615                lock, reply->lock_handle.cookie, *flags);
616
617         /* If enqueue returned a blocked lock but the completion handler has
618          * already run, then it fixed up the resource and we don't need to do it
619          * again. */
620         if ((*flags) & LDLM_FL_LOCK_CHANGED) {
621                 int newmode = reply->lock_desc.l_req_mode;
622                 LASSERT(!is_replay);
623                 if (newmode && newmode != lock->l_req_mode) {
624                         LDLM_DEBUG(lock, "server returned different mode %s",
625                                    ldlm_lockname[newmode]);
626                         lock->l_req_mode = newmode;
627                 }
628
629                 if (!ldlm_res_eq(&reply->lock_desc.l_resource.lr_name,
630                                  &lock->l_resource->lr_name)) {
631                         CDEBUG(D_INFO, "remote intent success, locking "DLDLMRES
632                                        " instead of "DLDLMRES"\n",
633                                PLDLMRES(&reply->lock_desc.l_resource),
634                                PLDLMRES(lock->l_resource));
635
636                         rc = ldlm_lock_change_resource(ns, lock,
637                                         &reply->lock_desc.l_resource.lr_name);
638                         if (rc || lock->l_resource == NULL)
639                                 GOTO(cleanup, rc = -ENOMEM);
640                         LDLM_DEBUG(lock, "client-side enqueue, new resource");
641                 }
642                 if (with_policy)
643                         if (!(type == LDLM_IBITS &&
644                               !(exp_connect_flags(exp) & OBD_CONNECT_IBITS)))
645                                 /* We assume lock type cannot change on server*/
646                                 ldlm_convert_policy_to_local(exp,
647                                                 lock->l_resource->lr_type,
648                                                 &reply->lock_desc.l_policy_data,
649                                                 &lock->l_policy_data);
650                 if (type != LDLM_PLAIN)
651                         LDLM_DEBUG(lock,"client-side enqueue, new policy data");
652         }
653
654         if ((*flags) & LDLM_FL_AST_SENT ||
655             /* Cancel extent locks as soon as possible on a liblustre client,
656              * because it cannot handle asynchronous ASTs robustly (see
657              * bug 7311). */
658             (LIBLUSTRE_CLIENT && type == LDLM_EXTENT)) {
659                 lock_res_and_lock(lock);
660                 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_BL_AST;
661                 unlock_res_and_lock(lock);
662                 LDLM_DEBUG(lock, "enqueue reply includes blocking AST");
663         }
664
665         /* If the lock has already been granted by a completion AST, don't
666          * clobber the LVB with an older one. */
667         if (lvb_len != 0) {
668                 /* We must lock or a racing completion might update lvb without
669                  * letting us know and we'll clobber the correct value.
670                  * Cannot unlock after the check either, a that still leaves
671                  * a tiny window for completion to get in */
672                 lock_res_and_lock(lock);
673                 if (lock->l_req_mode != lock->l_granted_mode)
674                         rc = ldlm_fill_lvb(lock, &req->rq_pill, RCL_SERVER,
675                                            lock->l_lvb_data, size);
676                 unlock_res_and_lock(lock);
677                 if (rc < 0) {
678                         cleanup_phase = 1;
679                         GOTO(cleanup, rc);
680                 }
681         }
682
683         if (!is_replay) {
684                 rc = ldlm_lock_enqueue(ns, &lock, NULL, flags);
685                 if (lock->l_completion_ast != NULL) {
686                         int err = lock->l_completion_ast(lock, *flags, NULL);
687                         if (!rc)
688                                 rc = err;
689                         if (rc)
690                                 cleanup_phase = 1;
691                 }
692         }
693
694         if (lvb_len && lvb != NULL) {
695                 /* Copy the LVB here, and not earlier, because the completion
696                  * AST (if any) can override what we got in the reply */
697                 memcpy(lvb, lock->l_lvb_data, lvb_len);
698         }
699
700         LDLM_DEBUG(lock, "client-side enqueue END");
701         EXIT;
702 cleanup:
703         if (cleanup_phase == 1 && rc)
704                 failed_lock_cleanup(ns, lock, mode);
705         /* Put lock 2 times, the second reference is held by ldlm_cli_enqueue */
706         LDLM_LOCK_PUT(lock);
707         LDLM_LOCK_RELEASE(lock);
708         return rc;
709 }
710 EXPORT_SYMBOL(ldlm_cli_enqueue_fini);
711
712 /**
713  * Estimate number of lock handles that would fit into request of given
714  * size.  PAGE_SIZE-512 is to allow TCP/IP and LNET headers to fit into
715  * a single page on the send/receive side. XXX: 512 should be changed to
716  * more adequate value.
717  */
718 static inline int ldlm_req_handles_avail(int req_size, int off)
719 {
720         int avail;
721
722         avail = min_t(int, LDLM_MAXREQSIZE, PAGE_CACHE_SIZE - 512) - req_size;
723         if (likely(avail >= 0))
724                 avail /= (int)sizeof(struct lustre_handle);
725         else
726                 avail = 0;
727         avail += LDLM_LOCKREQ_HANDLES - off;
728
729         return avail;
730 }
731
732 static inline int ldlm_capsule_handles_avail(struct req_capsule *pill,
733                                              enum req_location loc,
734                                              int off)
735 {
736         int size = req_capsule_msg_size(pill, loc);
737         return ldlm_req_handles_avail(size, off);
738 }
739
740 static inline int ldlm_format_handles_avail(struct obd_import *imp,
741                                             const struct req_format *fmt,
742                                             enum req_location loc, int off)
743 {
744         int size = req_capsule_fmt_size(imp->imp_msg_magic, fmt, loc);
745         return ldlm_req_handles_avail(size, off);
746 }
747
748 /**
749  * Cancel LRU locks and pack them into the enqueue request. Pack there the given
750  * \a count locks in \a cancels.
751  *
752  * This is to be called by functions preparing their own requests that
753  * might contain lists of locks to cancel in addition to actual operation
754  * that needs to be performed.
755  */
756 int ldlm_prep_elc_req(struct obd_export *exp, struct ptlrpc_request *req,
757                       int version, int opc, int canceloff,
758                       cfs_list_t *cancels, int count)
759 {
760         struct ldlm_namespace   *ns = exp->exp_obd->obd_namespace;
761         struct req_capsule      *pill = &req->rq_pill;
762         struct ldlm_request     *dlm = NULL;
763         int flags, avail, to_free, pack = 0;
764         CFS_LIST_HEAD(head);
765         int rc;
766         ENTRY;
767
768         if (cancels == NULL)
769                 cancels = &head;
770         if (ns_connect_cancelset(ns)) {
771                 /* Estimate the amount of available space in the request. */
772                 req_capsule_filled_sizes(pill, RCL_CLIENT);
773                 avail = ldlm_capsule_handles_avail(pill, RCL_CLIENT, canceloff);
774
775                 flags = ns_connect_lru_resize(ns) ?
776                         LDLM_CANCEL_LRUR : LDLM_CANCEL_AGED;
777                 to_free = !ns_connect_lru_resize(ns) &&
778                           opc == LDLM_ENQUEUE ? 1 : 0;
779
780                 /* Cancel LRU locks here _only_ if the server supports
781                  * EARLY_CANCEL. Otherwise we have to send extra CANCEL
782                  * RPC, which will make us slower. */
783                 if (avail > count)
784                         count += ldlm_cancel_lru_local(ns, cancels, to_free,
785                                                        avail - count, 0, flags);
786                 if (avail > count)
787                         pack = count;
788                 else
789                         pack = avail;
790                 req_capsule_set_size(pill, &RMF_DLM_REQ, RCL_CLIENT,
791                                      ldlm_request_bufsize(pack, opc));
792         }
793
794         rc = ptlrpc_request_pack(req, version, opc);
795         if (rc) {
796                 ldlm_lock_list_put(cancels, l_bl_ast, count);
797                 RETURN(rc);
798         }
799
800         if (ns_connect_cancelset(ns)) {
801                 if (canceloff) {
802                         dlm = req_capsule_client_get(pill, &RMF_DLM_REQ);
803                         LASSERT(dlm);
804                         /* Skip first lock handler in ldlm_request_pack(),
805                          * this method will incrment @lock_count according
806                          * to the lock handle amount actually written to
807                          * the buffer. */
808                         dlm->lock_count = canceloff;
809                 }
810                 /* Pack into the request @pack lock handles. */
811                 ldlm_cli_cancel_list(cancels, pack, req, 0);
812                 /* Prepare and send separate cancel RPC for others. */
813                 ldlm_cli_cancel_list(cancels, count - pack, NULL, 0);
814         } else {
815                 ldlm_lock_list_put(cancels, l_bl_ast, count);
816         }
817         RETURN(0);
818 }
819 EXPORT_SYMBOL(ldlm_prep_elc_req);
820
821 int ldlm_prep_enqueue_req(struct obd_export *exp, struct ptlrpc_request *req,
822                           cfs_list_t *cancels, int count)
823 {
824         return ldlm_prep_elc_req(exp, req, LUSTRE_DLM_VERSION, LDLM_ENQUEUE,
825                                  LDLM_ENQUEUE_CANCEL_OFF, cancels, count);
826 }
827 EXPORT_SYMBOL(ldlm_prep_enqueue_req);
828
829 struct ptlrpc_request *ldlm_enqueue_pack(struct obd_export *exp, int lvb_len)
830 {
831         struct ptlrpc_request *req;
832         int rc;
833         ENTRY;
834
835         req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_LDLM_ENQUEUE);
836         if (req == NULL)
837                 RETURN(ERR_PTR(-ENOMEM));
838
839         rc = ldlm_prep_enqueue_req(exp, req, NULL, 0);
840         if (rc) {
841                 ptlrpc_request_free(req);
842                 RETURN(ERR_PTR(rc));
843         }
844
845         req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER, lvb_len);
846         ptlrpc_request_set_replen(req);
847         RETURN(req);
848 }
849 EXPORT_SYMBOL(ldlm_enqueue_pack);
850
851 /**
852  * Client-side lock enqueue.
853  *
854  * If a request has some specific initialisation it is passed in \a reqp,
855  * otherwise it is created in ldlm_cli_enqueue.
856  *
857  * Supports sync and async requests, pass \a async flag accordingly. If a
858  * request was created in ldlm_cli_enqueue and it is the async request,
859  * pass it to the caller in \a reqp.
860  */
861 int ldlm_cli_enqueue(struct obd_export *exp, struct ptlrpc_request **reqp,
862                      struct ldlm_enqueue_info *einfo,
863                      const struct ldlm_res_id *res_id,
864                      ldlm_policy_data_t const *policy, __u64 *flags,
865                      void *lvb, __u32 lvb_len, enum lvb_type lvb_type,
866                      struct lustre_handle *lockh, int async)
867 {
868         struct ldlm_namespace *ns;
869         struct ldlm_lock      *lock;
870         struct ldlm_request   *body;
871         int                    is_replay = *flags & LDLM_FL_REPLAY;
872         int                    req_passed_in = 1;
873         int                    rc, err;
874         struct ptlrpc_request *req;
875         ENTRY;
876
877         LASSERT(exp != NULL);
878
879         ns = exp->exp_obd->obd_namespace;
880
881         /* If we're replaying this lock, just check some invariants.
882          * If we're creating a new lock, get everything all setup nice. */
883         if (is_replay) {
884                 lock = ldlm_handle2lock_long(lockh, 0);
885                 LASSERT(lock != NULL);
886                 LDLM_DEBUG(lock, "client-side enqueue START");
887                 LASSERT(exp == lock->l_conn_export);
888         } else {
889                 const struct ldlm_callback_suite cbs = {
890                         .lcs_completion = einfo->ei_cb_cp,
891                         .lcs_blocking   = einfo->ei_cb_bl,
892                         .lcs_glimpse    = einfo->ei_cb_gl
893                 };
894                 lock = ldlm_lock_create(ns, res_id, einfo->ei_type,
895                                         einfo->ei_mode, &cbs, einfo->ei_cbdata,
896                                         lvb_len, lvb_type);
897                 if (lock == NULL)
898                         RETURN(-ENOMEM);
899                 /* for the local lock, add the reference */
900                 ldlm_lock_addref_internal(lock, einfo->ei_mode);
901                 ldlm_lock2handle(lock, lockh);
902                 if (policy != NULL) {
903                         /* INODEBITS_INTEROP: If the server does not support
904                          * inodebits, we will request a plain lock in the
905                          * descriptor (ldlm_lock2desc() below) but use an
906                          * inodebits lock internally with both bits set.
907                          */
908                         if (einfo->ei_type == LDLM_IBITS &&
909                             !(exp_connect_flags(exp) &
910                               OBD_CONNECT_IBITS))
911                                 lock->l_policy_data.l_inodebits.bits =
912                                         MDS_INODELOCK_LOOKUP |
913                                         MDS_INODELOCK_UPDATE;
914                         else
915                                 lock->l_policy_data = *policy;
916                 }
917
918                 if (einfo->ei_type == LDLM_EXTENT) {
919                         /* extent lock without policy is a bug */
920                         if (policy == NULL)
921                                 LBUG();
922
923                         lock->l_req_extent = policy->l_extent;
924                 }
925                 LDLM_DEBUG(lock, "client-side enqueue START, flags %llx\n",
926                            *flags);
927         }
928
929         lock->l_conn_export = exp;
930         lock->l_export = NULL;
931         lock->l_blocking_ast = einfo->ei_cb_bl;
932         lock->l_flags |= (*flags & (LDLM_FL_NO_LRU | LDLM_FL_EXCL));
933
934         /* lock not sent to server yet */
935
936         if (reqp == NULL || *reqp == NULL) {
937                 req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
938                                                 &RQF_LDLM_ENQUEUE,
939                                                 LUSTRE_DLM_VERSION,
940                                                 LDLM_ENQUEUE);
941                 if (req == NULL) {
942                         failed_lock_cleanup(ns, lock, einfo->ei_mode);
943                         LDLM_LOCK_RELEASE(lock);
944                         RETURN(-ENOMEM);
945                 }
946                 req_passed_in = 0;
947                 if (reqp)
948                         *reqp = req;
949         } else {
950                 int len;
951
952                 req = *reqp;
953                 len = req_capsule_get_size(&req->rq_pill, &RMF_DLM_REQ,
954                                            RCL_CLIENT);
955                 LASSERTF(len >= sizeof(*body), "buflen[%d] = %d, not %d\n",
956                          DLM_LOCKREQ_OFF, len, (int)sizeof(*body));
957         }
958
959         /* Dump lock data into the request buffer */
960         body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
961         ldlm_lock2desc(lock, &body->lock_desc);
962         body->lock_flags = ldlm_flags_to_wire(*flags);
963         body->lock_handle[0] = *lockh;
964
965         /* Continue as normal. */
966         if (!req_passed_in) {
967                 if (lvb_len > 0)
968                         req_capsule_extend(&req->rq_pill,
969                                            &RQF_LDLM_ENQUEUE_LVB);
970                 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
971                                      lvb_len);
972                 ptlrpc_request_set_replen(req);
973         }
974
975         /*
976          * Liblustre client doesn't get extent locks, except for O_APPEND case
977          * where [0, OBD_OBJECT_EOF] lock is taken, or truncate, where
978          * [i_size, OBD_OBJECT_EOF] lock is taken.
979          */
980         LASSERT(ergo(LIBLUSTRE_CLIENT, einfo->ei_type != LDLM_EXTENT ||
981                      policy->l_extent.end == OBD_OBJECT_EOF));
982
983         if (async) {
984                 LASSERT(reqp != NULL);
985                 RETURN(0);
986         }
987
988         LDLM_DEBUG(lock, "sending request");
989
990         rc = ptlrpc_queue_wait(req);
991
992         err = ldlm_cli_enqueue_fini(exp, req, einfo->ei_type, policy ? 1 : 0,
993                                     einfo->ei_mode, flags, lvb, lvb_len,
994                                     lockh, rc);
995
996         /* If ldlm_cli_enqueue_fini did not find the lock, we need to free
997          * one reference that we took */
998         if (err == -ENOLCK)
999                 LDLM_LOCK_RELEASE(lock);
1000         else
1001                 rc = err;
1002
1003         if (!req_passed_in && req != NULL) {
1004                 ptlrpc_req_finished(req);
1005                 if (reqp)
1006                         *reqp = NULL;
1007         }
1008
1009         RETURN(rc);
1010 }
1011 EXPORT_SYMBOL(ldlm_cli_enqueue);
1012
1013 static int ldlm_cli_convert_local(struct ldlm_lock *lock, int new_mode,
1014                                   __u32 *flags)
1015 {
1016         struct ldlm_resource *res;
1017         int rc;
1018         ENTRY;
1019         if (ns_is_client(ldlm_lock_to_ns(lock))) {
1020                 CERROR("Trying to cancel local lock\n");
1021                 LBUG();
1022         }
1023         LDLM_DEBUG(lock, "client-side local convert");
1024
1025         res = ldlm_lock_convert(lock, new_mode, flags);
1026         if (res) {
1027                 ldlm_reprocess_all(res);
1028                 rc = 0;
1029         } else {
1030                 rc = LUSTRE_EDEADLK;
1031         }
1032         LDLM_DEBUG(lock, "client-side local convert handler END");
1033         LDLM_LOCK_PUT(lock);
1034         RETURN(rc);
1035 }
1036
1037 /* FIXME: one of ldlm_cli_convert or the server side should reject attempted
1038  * conversion of locks which are on the waiting or converting queue */
1039 /* Caller of this code is supposed to take care of lock readers/writers
1040    accounting */
1041 int ldlm_cli_convert(struct lustre_handle *lockh, int new_mode, __u32 *flags)
1042 {
1043         struct ldlm_request   *body;
1044         struct ldlm_reply     *reply;
1045         struct ldlm_lock      *lock;
1046         struct ldlm_resource  *res;
1047         struct ptlrpc_request *req;
1048         int                    rc;
1049         ENTRY;
1050
1051         lock = ldlm_handle2lock(lockh);
1052         if (!lock) {
1053                 LBUG();
1054                 RETURN(-EINVAL);
1055         }
1056         *flags = 0;
1057
1058         if (lock->l_conn_export == NULL)
1059                 RETURN(ldlm_cli_convert_local(lock, new_mode, flags));
1060
1061         LDLM_DEBUG(lock, "client-side convert");
1062
1063         req = ptlrpc_request_alloc_pack(class_exp2cliimp(lock->l_conn_export),
1064                                         &RQF_LDLM_CONVERT, LUSTRE_DLM_VERSION,
1065                                         LDLM_CONVERT);
1066         if (req == NULL) {
1067                 LDLM_LOCK_PUT(lock);
1068                 RETURN(-ENOMEM);
1069         }
1070
1071         body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1072         body->lock_handle[0] = lock->l_remote_handle;
1073
1074         body->lock_desc.l_req_mode = new_mode;
1075         body->lock_flags = ldlm_flags_to_wire(*flags);
1076
1077
1078         ptlrpc_request_set_replen(req);
1079         rc = ptlrpc_queue_wait(req);
1080         if (rc != ELDLM_OK)
1081                 GOTO(out, rc);
1082
1083         reply = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
1084         if (reply == NULL)
1085                 GOTO(out, rc = -EPROTO);
1086
1087         if (req->rq_status)
1088                 GOTO(out, rc = req->rq_status);
1089
1090         res = ldlm_lock_convert(lock, new_mode, &reply->lock_flags);
1091         if (res != NULL) {
1092                 ldlm_reprocess_all(res);
1093                 /* Go to sleep until the lock is granted. */
1094                 /* FIXME: or cancelled. */
1095                 if (lock->l_completion_ast) {
1096                         rc = lock->l_completion_ast(lock, LDLM_FL_WAIT_NOREPROC,
1097                                                     NULL);
1098                         if (rc)
1099                                 GOTO(out, rc);
1100                 }
1101         } else {
1102                 rc = LUSTRE_EDEADLK;
1103         }
1104         EXIT;
1105  out:
1106         LDLM_LOCK_PUT(lock);
1107         ptlrpc_req_finished(req);
1108         return rc;
1109 }
1110 EXPORT_SYMBOL(ldlm_cli_convert);
1111
1112 /**
1113  * Cancel locks locally.
1114  * Returns:
1115  * \retval LDLM_FL_LOCAL_ONLY if there is no need for a CANCEL RPC to the server
1116  * \retval LDLM_FL_CANCELING otherwise;
1117  * \retval LDLM_FL_BL_AST if there is a need for a separate CANCEL RPC.
1118  */
1119 static __u64 ldlm_cli_cancel_local(struct ldlm_lock *lock)
1120 {
1121         __u64 rc = LDLM_FL_LOCAL_ONLY;
1122         ENTRY;
1123
1124         if (lock->l_conn_export) {
1125                 bool local_only;
1126
1127                 LDLM_DEBUG(lock, "client-side cancel");
1128                 /* Set this flag to prevent others from getting new references*/
1129                 lock_res_and_lock(lock);
1130                 ldlm_set_cbpending(lock);
1131                 local_only = !!(lock->l_flags &
1132                                 (LDLM_FL_LOCAL_ONLY|LDLM_FL_CANCEL_ON_BLOCK));
1133                 ldlm_cancel_callback(lock);
1134                 rc = (ldlm_is_bl_ast(lock)) ?
1135                         LDLM_FL_BL_AST : LDLM_FL_CANCELING;
1136                 unlock_res_and_lock(lock);
1137
1138                 if (local_only) {
1139                         CDEBUG(D_DLMTRACE, "not sending request (at caller's "
1140                                "instruction)\n");
1141                         rc = LDLM_FL_LOCAL_ONLY;
1142                 }
1143                 ldlm_lock_cancel(lock);
1144         } else {
1145                 if (ns_is_client(ldlm_lock_to_ns(lock))) {
1146                         LDLM_ERROR(lock, "Trying to cancel local lock");
1147                         LBUG();
1148                 }
1149                 LDLM_DEBUG(lock, "server-side local cancel");
1150                 ldlm_lock_cancel(lock);
1151                 ldlm_reprocess_all(lock->l_resource);
1152         }
1153
1154         RETURN(rc);
1155 }
1156
1157 /**
1158  * Pack \a count locks in \a head into ldlm_request buffer of request \a req.
1159  */
1160 static void ldlm_cancel_pack(struct ptlrpc_request *req,
1161                              cfs_list_t *head, int count)
1162 {
1163         struct ldlm_request *dlm;
1164         struct ldlm_lock *lock;
1165         int max, packed = 0;
1166         ENTRY;
1167
1168         dlm = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1169         LASSERT(dlm != NULL);
1170
1171         /* Check the room in the request buffer. */
1172         max = req_capsule_get_size(&req->rq_pill, &RMF_DLM_REQ, RCL_CLIENT) -
1173                 sizeof(struct ldlm_request);
1174         max /= sizeof(struct lustre_handle);
1175         max += LDLM_LOCKREQ_HANDLES;
1176         LASSERT(max >= dlm->lock_count + count);
1177
1178         /* XXX: it would be better to pack lock handles grouped by resource.
1179          * so that the server cancel would call filter_lvbo_update() less
1180          * frequently. */
1181         cfs_list_for_each_entry(lock, head, l_bl_ast) {
1182                 if (!count--)
1183                         break;
1184                 LASSERT(lock->l_conn_export);
1185                 /* Pack the lock handle to the given request buffer. */
1186                 LDLM_DEBUG(lock, "packing");
1187                 dlm->lock_handle[dlm->lock_count++] = lock->l_remote_handle;
1188                 packed++;
1189         }
1190         CDEBUG(D_DLMTRACE, "%d locks packed\n", packed);
1191         EXIT;
1192 }
1193
1194 /**
1195  * Prepare and send a batched cancel RPC. It will include \a count lock
1196  * handles of locks given in \a cancels list. */
1197 int ldlm_cli_cancel_req(struct obd_export *exp, cfs_list_t *cancels,
1198                         int count, ldlm_cancel_flags_t flags)
1199 {
1200         struct ptlrpc_request *req = NULL;
1201         struct obd_import *imp;
1202         int free, sent = 0;
1203         int rc = 0;
1204         ENTRY;
1205
1206         LASSERT(exp != NULL);
1207         LASSERT(count > 0);
1208
1209         CFS_FAIL_TIMEOUT(OBD_FAIL_LDLM_PAUSE_CANCEL, cfs_fail_val);
1210
1211         if (CFS_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_RACE))
1212                 RETURN(count);
1213
1214         free = ldlm_format_handles_avail(class_exp2cliimp(exp),
1215                                          &RQF_LDLM_CANCEL, RCL_CLIENT, 0);
1216         if (count > free)
1217                 count = free;
1218
1219         while (1) {
1220                 imp = class_exp2cliimp(exp);
1221                 if (imp == NULL || imp->imp_invalid) {
1222                         CDEBUG(D_DLMTRACE,
1223                                "skipping cancel on invalid import %p\n", imp);
1224                         RETURN(count);
1225                 }
1226
1227                 req = ptlrpc_request_alloc(imp, &RQF_LDLM_CANCEL);
1228                 if (req == NULL)
1229                         GOTO(out, rc = -ENOMEM);
1230
1231                 req_capsule_filled_sizes(&req->rq_pill, RCL_CLIENT);
1232                 req_capsule_set_size(&req->rq_pill, &RMF_DLM_REQ, RCL_CLIENT,
1233                                      ldlm_request_bufsize(count, LDLM_CANCEL));
1234
1235                 rc = ptlrpc_request_pack(req, LUSTRE_DLM_VERSION, LDLM_CANCEL);
1236                 if (rc) {
1237                         ptlrpc_request_free(req);
1238                         GOTO(out, rc);
1239                 }
1240
1241                 req->rq_request_portal = LDLM_CANCEL_REQUEST_PORTAL;
1242                 req->rq_reply_portal = LDLM_CANCEL_REPLY_PORTAL;
1243                 ptlrpc_at_set_req_timeout(req);
1244
1245                 ldlm_cancel_pack(req, cancels, count);
1246
1247                 ptlrpc_request_set_replen(req);
1248                 if (flags & LCF_ASYNC) {
1249                         ptlrpcd_add_req(req, PDL_POLICY_LOCAL, -1);
1250                         sent = count;
1251                         GOTO(out, 0);
1252                 } else {
1253                         rc = ptlrpc_queue_wait(req);
1254                 }
1255                 if (rc == LUSTRE_ESTALE) {
1256                         CDEBUG(D_DLMTRACE, "client/server (nid %s) "
1257                                "out of sync -- not fatal\n",
1258                                libcfs_nid2str(req->rq_import->
1259                                               imp_connection->c_peer.nid));
1260                         rc = 0;
1261                 } else if (rc == -ETIMEDOUT && /* check there was no reconnect*/
1262                            req->rq_import_generation == imp->imp_generation) {
1263                         ptlrpc_req_finished(req);
1264                         continue;
1265                 } else if (rc != ELDLM_OK) {
1266                         /* -ESHUTDOWN is common on umount */
1267                         CDEBUG_LIMIT(rc == -ESHUTDOWN ? D_DLMTRACE : D_ERROR,
1268                                      "Got rc %d from cancel RPC: "
1269                                      "canceling anyway\n", rc);
1270                         break;
1271                 }
1272                 sent = count;
1273                 break;
1274         }
1275
1276         ptlrpc_req_finished(req);
1277         EXIT;
1278 out:
1279         return sent ? sent : rc;
1280 }
1281 EXPORT_SYMBOL(ldlm_cli_cancel_req);
1282
1283 static inline struct ldlm_pool *ldlm_imp2pl(struct obd_import *imp)
1284 {
1285         LASSERT(imp != NULL);
1286         return &imp->imp_obd->obd_namespace->ns_pool;
1287 }
1288
1289 /**
1290  * Update client's OBD pool related fields with new SLV and Limit from \a req.
1291  */
1292 int ldlm_cli_update_pool(struct ptlrpc_request *req)
1293 {
1294         struct obd_device *obd;
1295         __u64 new_slv;
1296         __u32 new_limit;
1297         ENTRY;
1298         if (unlikely(!req->rq_import || !req->rq_import->imp_obd ||
1299                      !imp_connect_lru_resize(req->rq_import)))
1300         {
1301                 /*
1302                  * Do nothing for corner cases.
1303                  */
1304                 RETURN(0);
1305         }
1306
1307         /* In some cases RPC may contain SLV and limit zeroed out. This
1308          * is the case when server does not support LRU resize feature.
1309          * This is also possible in some recovery cases when server-side
1310          * reqs have no reference to the OBD export and thus access to
1311          * server-side namespace is not possible. */
1312         if (lustre_msg_get_slv(req->rq_repmsg) == 0 ||
1313             lustre_msg_get_limit(req->rq_repmsg) == 0) {
1314                 DEBUG_REQ(D_HA, req, "Zero SLV or Limit found "
1315                           "(SLV: "LPU64", Limit: %u)",
1316                           lustre_msg_get_slv(req->rq_repmsg),
1317                           lustre_msg_get_limit(req->rq_repmsg));
1318                 RETURN(0);
1319         }
1320
1321         new_limit = lustre_msg_get_limit(req->rq_repmsg);
1322         new_slv = lustre_msg_get_slv(req->rq_repmsg);
1323         obd = req->rq_import->imp_obd;
1324
1325         /* Set new SLV and limit in OBD fields to make them accessible
1326          * to the pool thread. We do not access obd_namespace and pool
1327          * directly here as there is no reliable way to make sure that
1328          * they are still alive at cleanup time. Evil races are possible
1329          * which may cause Oops at that time. */
1330         write_lock(&obd->obd_pool_lock);
1331         obd->obd_pool_slv = new_slv;
1332         obd->obd_pool_limit = new_limit;
1333         write_unlock(&obd->obd_pool_lock);
1334
1335         RETURN(0);
1336 }
1337 EXPORT_SYMBOL(ldlm_cli_update_pool);
1338
1339 /**
1340  * Client side lock cancel.
1341  *
1342  * Lock must not have any readers or writers by this time.
1343  */
1344 int ldlm_cli_cancel(struct lustre_handle *lockh,
1345                     ldlm_cancel_flags_t cancel_flags)
1346 {
1347         struct obd_export *exp;
1348         int avail, flags, count = 1;
1349         __u64 rc = 0;
1350         struct ldlm_namespace *ns;
1351         struct ldlm_lock *lock;
1352         CFS_LIST_HEAD(cancels);
1353         ENTRY;
1354
1355         /* concurrent cancels on the same handle can happen */
1356         lock = ldlm_handle2lock_long(lockh, LDLM_FL_CANCELING);
1357         if (lock == NULL) {
1358                 LDLM_DEBUG_NOLOCK("lock is already being destroyed\n");
1359                 RETURN(0);
1360         }
1361
1362         rc = ldlm_cli_cancel_local(lock);
1363         if (rc == LDLM_FL_LOCAL_ONLY || cancel_flags & LCF_LOCAL) {
1364                 LDLM_LOCK_RELEASE(lock);
1365                 RETURN(0);
1366         }
1367         /* Even if the lock is marked as LDLM_FL_BL_AST, this is a LDLM_CANCEL
1368          * RPC which goes to canceld portal, so we can cancel other LRU locks
1369          * here and send them all as one LDLM_CANCEL RPC. */
1370         LASSERT(cfs_list_empty(&lock->l_bl_ast));
1371         cfs_list_add(&lock->l_bl_ast, &cancels);
1372
1373         exp = lock->l_conn_export;
1374         if (exp_connect_cancelset(exp)) {
1375                 avail = ldlm_format_handles_avail(class_exp2cliimp(exp),
1376                                                   &RQF_LDLM_CANCEL,
1377                                                   RCL_CLIENT, 0);
1378                 LASSERT(avail > 0);
1379
1380                 ns = ldlm_lock_to_ns(lock);
1381                 flags = ns_connect_lru_resize(ns) ?
1382                         LDLM_CANCEL_LRUR : LDLM_CANCEL_AGED;
1383                 count += ldlm_cancel_lru_local(ns, &cancels, 0, avail - 1,
1384                                                LCF_BL_AST, flags);
1385         }
1386         ldlm_cli_cancel_list(&cancels, count, NULL, cancel_flags);
1387         RETURN(0);
1388 }
1389 EXPORT_SYMBOL(ldlm_cli_cancel);
1390
1391 /**
1392  * Locally cancel up to \a count locks in list \a cancels.
1393  * Return the number of cancelled locks.
1394  */
1395 int ldlm_cli_cancel_list_local(cfs_list_t *cancels, int count,
1396                                ldlm_cancel_flags_t flags)
1397 {
1398         CFS_LIST_HEAD(head);
1399         struct ldlm_lock *lock, *next;
1400         int left = 0, bl_ast = 0;
1401         __u64 rc;
1402
1403         left = count;
1404         cfs_list_for_each_entry_safe(lock, next, cancels, l_bl_ast) {
1405                 if (left-- == 0)
1406                         break;
1407
1408                 if (flags & LCF_LOCAL) {
1409                         rc = LDLM_FL_LOCAL_ONLY;
1410                         ldlm_lock_cancel(lock);
1411                 } else {
1412                         rc = ldlm_cli_cancel_local(lock);
1413                 }
1414                 /* Until we have compound requests and can send LDLM_CANCEL
1415                  * requests batched with generic RPCs, we need to send cancels
1416                  * with the LDLM_FL_BL_AST flag in a separate RPC from
1417                  * the one being generated now. */
1418                 if (!(flags & LCF_BL_AST) && (rc == LDLM_FL_BL_AST)) {
1419                         LDLM_DEBUG(lock, "Cancel lock separately");
1420                         cfs_list_del_init(&lock->l_bl_ast);
1421                         cfs_list_add(&lock->l_bl_ast, &head);
1422                         bl_ast++;
1423                         continue;
1424                 }
1425                 if (rc == LDLM_FL_LOCAL_ONLY) {
1426                         /* CANCEL RPC should not be sent to server. */
1427                         cfs_list_del_init(&lock->l_bl_ast);
1428                         LDLM_LOCK_RELEASE(lock);
1429                         count--;
1430                 }
1431         }
1432         if (bl_ast > 0) {
1433                 count -= bl_ast;
1434                 ldlm_cli_cancel_list(&head, bl_ast, NULL, 0);
1435         }
1436
1437         RETURN(count);
1438 }
1439 EXPORT_SYMBOL(ldlm_cli_cancel_list_local);
1440
1441 /**
1442  * Cancel as many locks as possible w/o sending any RPCs (e.g. to write back
1443  * dirty data, to close a file, ...) or waiting for any RPCs in-flight (e.g.
1444  * readahead requests, ...)
1445  */
1446 static ldlm_policy_res_t ldlm_cancel_no_wait_policy(struct ldlm_namespace *ns,
1447                                                     struct ldlm_lock *lock,
1448                                                     int unused, int added,
1449                                                     int count)
1450 {
1451         ldlm_policy_res_t result = LDLM_POLICY_CANCEL_LOCK;
1452         ldlm_cancel_for_recovery cb = ns->ns_cancel_for_recovery;
1453
1454         /* don't check added & count since we want to process all locks
1455          * from unused list.
1456          * It's fine to not take lock to access lock->l_resource since
1457          * the lock has already been granted so it won't change. */
1458         switch (lock->l_resource->lr_type) {
1459                 case LDLM_EXTENT:
1460                 case LDLM_IBITS:
1461                         if (cb && cb(lock))
1462                                 break;
1463                 default:
1464                         result = LDLM_POLICY_SKIP_LOCK;
1465                         lock_res_and_lock(lock);
1466                         ldlm_set_skipped(lock);
1467                         unlock_res_and_lock(lock);
1468                         break;
1469         }
1470
1471         RETURN(result);
1472 }
1473
1474 /**
1475  * Callback function for LRU-resize policy. Decides whether to keep
1476  * \a lock in LRU for current \a LRU size \a unused, added in current
1477  * scan \a added and number of locks to be preferably canceled \a count.
1478  *
1479  * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1480  *
1481  * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1482  */
1483 static ldlm_policy_res_t ldlm_cancel_lrur_policy(struct ldlm_namespace *ns,
1484                                                  struct ldlm_lock *lock,
1485                                                  int unused, int added,
1486                                                  int count)
1487 {
1488         cfs_time_t cur = cfs_time_current();
1489         struct ldlm_pool *pl = &ns->ns_pool;
1490         __u64 slv, lvf, lv;
1491         cfs_time_t la;
1492
1493         /* Stop LRU processing when we reach past @count or have checked all
1494          * locks in LRU. */
1495         if (count && added >= count)
1496                 return LDLM_POLICY_KEEP_LOCK;
1497
1498         slv = ldlm_pool_get_slv(pl);
1499         lvf = ldlm_pool_get_lvf(pl);
1500         la = cfs_duration_sec(cfs_time_sub(cur,
1501                               lock->l_last_used));
1502         lv = lvf * la * unused;
1503
1504         /* Inform pool about current CLV to see it via proc. */
1505         ldlm_pool_set_clv(pl, lv);
1506
1507         /* Stop when SLV is not yet come from server or lv is smaller than
1508          * it is. */
1509         return (slv == 0 || lv < slv) ?
1510                 LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK;
1511 }
1512
1513 /**
1514  * Callback function for proc used policy. Makes decision whether to keep
1515  * \a lock in LRU for current \a LRU size \a unused, added in current scan \a
1516  * added and number of locks to be preferably canceled \a count.
1517  *
1518  * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1519  *
1520  * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1521  */
1522 static ldlm_policy_res_t ldlm_cancel_passed_policy(struct ldlm_namespace *ns,
1523                                                    struct ldlm_lock *lock,
1524                                                    int unused, int added,
1525                                                    int count)
1526 {
1527         /* Stop LRU processing when we reach past @count or have checked all
1528          * locks in LRU. */
1529         return (added >= count) ?
1530                 LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK;
1531 }
1532
1533 /**
1534  * Callback function for aged policy. Makes decision whether to keep \a lock in
1535  * LRU for current LRU size \a unused, added in current scan \a added and
1536  * number of locks to be preferably canceled \a count.
1537  *
1538  * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1539  *
1540  * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1541  */
1542 static ldlm_policy_res_t ldlm_cancel_aged_policy(struct ldlm_namespace *ns,
1543                                                  struct ldlm_lock *lock,
1544                                                  int unused, int added,
1545                                                  int count)
1546 {
1547         /* Stop LRU processing if young lock is found and we reach past count */
1548         return ((added >= count) &&
1549                 cfs_time_before(cfs_time_current(),
1550                                 cfs_time_add(lock->l_last_used,
1551                                              ns->ns_max_age))) ?
1552                 LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK;
1553 }
1554
1555 /**
1556  * Callback function for default policy. Makes decision whether to keep \a lock
1557  * in LRU for current LRU size \a unused, added in current scan \a added and
1558  * number of locks to be preferably canceled \a count.
1559  *
1560  * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1561  *
1562  * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1563  */
1564 static ldlm_policy_res_t ldlm_cancel_default_policy(struct ldlm_namespace *ns,
1565                                                     struct ldlm_lock *lock,
1566                                                     int unused, int added,
1567                                                     int count)
1568 {
1569         /* Stop LRU processing when we reach past count or have checked all
1570          * locks in LRU. */
1571         return (added >= count) ?
1572                 LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK;
1573 }
1574
1575 typedef ldlm_policy_res_t (*ldlm_cancel_lru_policy_t)(struct ldlm_namespace *,
1576                                                       struct ldlm_lock *, int,
1577                                                       int, int);
1578
1579 static ldlm_cancel_lru_policy_t
1580 ldlm_cancel_lru_policy(struct ldlm_namespace *ns, int flags)
1581 {
1582         if (flags & LDLM_CANCEL_NO_WAIT)
1583                 return ldlm_cancel_no_wait_policy;
1584
1585         if (ns_connect_lru_resize(ns)) {
1586                 if (flags & LDLM_CANCEL_SHRINK)
1587                         /* We kill passed number of old locks. */
1588                         return ldlm_cancel_passed_policy;
1589                 else if (flags & LDLM_CANCEL_LRUR)
1590                         return ldlm_cancel_lrur_policy;
1591                 else if (flags & LDLM_CANCEL_PASSED)
1592                         return ldlm_cancel_passed_policy;
1593         } else {
1594                 if (flags & LDLM_CANCEL_AGED)
1595                         return ldlm_cancel_aged_policy;
1596         }
1597
1598         return ldlm_cancel_default_policy;
1599 }
1600
1601 /**
1602  * - Free space in LRU for \a count new locks,
1603  *   redundant unused locks are canceled locally;
1604  * - also cancel locally unused aged locks;
1605  * - do not cancel more than \a max locks;
1606  * - GET the found locks and add them into the \a cancels list.
1607  *
1608  * A client lock can be added to the l_bl_ast list only when it is
1609  * marked LDLM_FL_CANCELING. Otherwise, somebody is already doing
1610  * CANCEL.  There are the following use cases:
1611  * ldlm_cancel_resource_local(), ldlm_cancel_lru_local() and
1612  * ldlm_cli_cancel(), which check and set this flag properly. As any
1613  * attempt to cancel a lock rely on this flag, l_bl_ast list is accessed
1614  * later without any special locking.
1615  *
1616  * Calling policies for enabled LRU resize:
1617  * ----------------------------------------
1618  * flags & LDLM_CANCEL_LRUR - use LRU resize policy (SLV from server) to
1619  *                            cancel not more than \a count locks;
1620  *
1621  * flags & LDLM_CANCEL_PASSED - cancel \a count number of old locks (located at
1622  *                              the beginning of LRU list);
1623  *
1624  * flags & LDLM_CANCEL_SHRINK - cancel not more than \a count locks according to
1625  *                              memory pressre policy function;
1626  *
1627  * flags & LDLM_CANCEL_AGED - cancel \a count locks according to "aged policy".
1628  *
1629  * flags & LDLM_CANCEL_NO_WAIT - cancel as many unused locks as possible
1630  *                               (typically before replaying locks) w/o
1631  *                               sending any RPCs or waiting for any
1632  *                               outstanding RPC to complete.
1633  */
1634 static int ldlm_prepare_lru_list(struct ldlm_namespace *ns, cfs_list_t *cancels,
1635                                  int count, int max, int flags)
1636 {
1637         ldlm_cancel_lru_policy_t pf;
1638         struct ldlm_lock *lock, *next;
1639         int added = 0, unused, remained;
1640         ENTRY;
1641
1642         spin_lock(&ns->ns_lock);
1643         unused = ns->ns_nr_unused;
1644         remained = unused;
1645
1646         if (!ns_connect_lru_resize(ns))
1647                 count += unused - ns->ns_max_unused;
1648
1649         pf = ldlm_cancel_lru_policy(ns, flags);
1650         LASSERT(pf != NULL);
1651
1652         while (!cfs_list_empty(&ns->ns_unused_list)) {
1653                 ldlm_policy_res_t result;
1654
1655                 /* all unused locks */
1656                 if (remained-- <= 0)
1657                         break;
1658
1659                 /* For any flags, stop scanning if @max is reached. */
1660                 if (max && added >= max)
1661                         break;
1662
1663                 cfs_list_for_each_entry_safe(lock, next, &ns->ns_unused_list,
1664                                              l_lru) {
1665                         /* No locks which got blocking requests. */
1666                         LASSERT(!ldlm_is_bl_ast(lock));
1667
1668                         if (flags & LDLM_CANCEL_NO_WAIT &&
1669                             ldlm_is_skipped(lock))
1670                                 /* already processed */
1671                                 continue;
1672
1673                         /* Somebody is already doing CANCEL. No need for this
1674                          * lock in LRU, do not traverse it again. */
1675                         if (!ldlm_is_canceling(lock))
1676                                 break;
1677
1678                         ldlm_lock_remove_from_lru_nolock(lock);
1679                 }
1680                 if (&lock->l_lru == &ns->ns_unused_list)
1681                         break;
1682
1683                 LDLM_LOCK_GET(lock);
1684                 spin_unlock(&ns->ns_lock);
1685                 lu_ref_add(&lock->l_reference, __FUNCTION__, current);
1686
1687                 /* Pass the lock through the policy filter and see if it
1688                  * should stay in LRU.
1689                  *
1690                  * Even for shrinker policy we stop scanning if
1691                  * we find a lock that should stay in the cache.
1692                  * We should take into account lock age anyway
1693                  * as a new lock is a valuable resource even if
1694                  * it has a low weight.
1695                  *
1696                  * That is, for shrinker policy we drop only
1697                  * old locks, but additionally choose them by
1698                  * their weight. Big extent locks will stay in
1699                  * the cache. */
1700                 result = pf(ns, lock, unused, added, count);
1701                 if (result == LDLM_POLICY_KEEP_LOCK) {
1702                         lu_ref_del(&lock->l_reference,
1703                                    __FUNCTION__, current);
1704                         LDLM_LOCK_RELEASE(lock);
1705                         spin_lock(&ns->ns_lock);
1706                         break;
1707                 }
1708                 if (result == LDLM_POLICY_SKIP_LOCK) {
1709                         lu_ref_del(&lock->l_reference,
1710                                    __func__, current);
1711                         LDLM_LOCK_RELEASE(lock);
1712                         spin_lock(&ns->ns_lock);
1713                         continue;
1714                 }
1715
1716                 lock_res_and_lock(lock);
1717                 /* Check flags again under the lock. */
1718                 if (ldlm_is_canceling(lock) ||
1719                     (ldlm_lock_remove_from_lru(lock) == 0)) {
1720                         /* Another thread is removing lock from LRU, or
1721                          * somebody is already doing CANCEL, or there
1722                          * is a blocking request which will send cancel
1723                          * by itself, or the lock is no longer unused. */
1724                         unlock_res_and_lock(lock);
1725                         lu_ref_del(&lock->l_reference, __FUNCTION__, current);
1726                         LDLM_LOCK_RELEASE(lock);
1727                         spin_lock(&ns->ns_lock);
1728                         continue;
1729                 }
1730                 LASSERT(!lock->l_readers && !lock->l_writers);
1731
1732                 /* If we have chosen to cancel this lock voluntarily, we
1733                  * better send cancel notification to server, so that it
1734                  * frees appropriate state. This might lead to a race
1735                  * where while we are doing cancel here, server is also
1736                  * silently cancelling this lock. */
1737                 ldlm_clear_cancel_on_block(lock);
1738
1739                 /* Setting the CBPENDING flag is a little misleading,
1740                  * but prevents an important race; namely, once
1741                  * CBPENDING is set, the lock can accumulate no more
1742                  * readers/writers. Since readers and writers are
1743                  * already zero here, ldlm_lock_decref() won't see
1744                  * this flag and call l_blocking_ast */
1745                 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_CANCELING;
1746
1747                 /* We can't re-add to l_lru as it confuses the
1748                  * refcounting in ldlm_lock_remove_from_lru() if an AST
1749                  * arrives after we drop lr_lock below. We use l_bl_ast
1750                  * and can't use l_pending_chain as it is used both on
1751                  * server and client nevertheless bug 5666 says it is
1752                  * used only on server */
1753                 LASSERT(cfs_list_empty(&lock->l_bl_ast));
1754                 cfs_list_add(&lock->l_bl_ast, cancels);
1755                 unlock_res_and_lock(lock);
1756                 lu_ref_del(&lock->l_reference, __FUNCTION__, current);
1757                 spin_lock(&ns->ns_lock);
1758                 added++;
1759                 unused--;
1760         }
1761         spin_unlock(&ns->ns_lock);
1762         RETURN(added);
1763 }
1764
1765 int ldlm_cancel_lru_local(struct ldlm_namespace *ns, cfs_list_t *cancels,
1766                           int count, int max, ldlm_cancel_flags_t cancel_flags,
1767                           int flags)
1768 {
1769         int added;
1770         added = ldlm_prepare_lru_list(ns, cancels, count, max, flags);
1771         if (added <= 0)
1772                 return added;
1773         return ldlm_cli_cancel_list_local(cancels, added, cancel_flags);
1774 }
1775
1776 /**
1777  * Cancel at least \a nr locks from given namespace LRU.
1778  *
1779  * When called with LCF_ASYNC the blocking callback will be handled
1780  * in a thread and this function will return after the thread has been
1781  * asked to call the callback.  When called with LCF_ASYNC the blocking
1782  * callback will be performed in this function.
1783  */
1784 int ldlm_cancel_lru(struct ldlm_namespace *ns, int nr,
1785                     ldlm_cancel_flags_t cancel_flags,
1786                     int flags)
1787 {
1788         CFS_LIST_HEAD(cancels);
1789         int count, rc;
1790         ENTRY;
1791
1792 #ifndef __KERNEL__
1793         cancel_flags &= ~LCF_ASYNC; /* force to be sync in user space */
1794 #endif
1795         /* Just prepare the list of locks, do not actually cancel them yet.
1796          * Locks are cancelled later in a separate thread. */
1797         count = ldlm_prepare_lru_list(ns, &cancels, nr, 0, flags);
1798         rc = ldlm_bl_to_thread_list(ns, NULL, &cancels, count, cancel_flags);
1799         if (rc == 0)
1800                 RETURN(count);
1801
1802         RETURN(0);
1803 }
1804
1805 /**
1806  * Find and cancel locally unused locks found on resource, matched to the
1807  * given policy, mode. GET the found locks and add them into the \a cancels
1808  * list.
1809  */
1810 int ldlm_cancel_resource_local(struct ldlm_resource *res,
1811                                cfs_list_t *cancels,
1812                                ldlm_policy_data_t *policy,
1813                                ldlm_mode_t mode, __u64 lock_flags,
1814                                ldlm_cancel_flags_t cancel_flags, void *opaque)
1815 {
1816         struct ldlm_lock *lock;
1817         int count = 0;
1818         ENTRY;
1819
1820         lock_res(res);
1821         cfs_list_for_each_entry(lock, &res->lr_granted, l_res_link) {
1822                 if (opaque != NULL && lock->l_ast_data != opaque) {
1823                         LDLM_ERROR(lock, "data %p doesn't match opaque %p",
1824                                    lock->l_ast_data, opaque);
1825                         //LBUG();
1826                         continue;
1827                 }
1828
1829                 if (lock->l_readers || lock->l_writers)
1830                         continue;
1831
1832                 /* If somebody is already doing CANCEL, or blocking AST came,
1833                  * skip this lock. */
1834                 if (ldlm_is_bl_ast(lock) || ldlm_is_canceling(lock))
1835                         continue;
1836
1837                 if (lockmode_compat(lock->l_granted_mode, mode))
1838                         continue;
1839
1840                 /* If policy is given and this is IBITS lock, add to list only
1841                  * those locks that match by policy. */
1842                 if (policy && (lock->l_resource->lr_type == LDLM_IBITS) &&
1843                     !(lock->l_policy_data.l_inodebits.bits &
1844                       policy->l_inodebits.bits))
1845                         continue;
1846
1847                 /* See CBPENDING comment in ldlm_cancel_lru */
1848                 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_CANCELING |
1849                                  lock_flags;
1850
1851                 LASSERT(cfs_list_empty(&lock->l_bl_ast));
1852                 cfs_list_add(&lock->l_bl_ast, cancels);
1853                 LDLM_LOCK_GET(lock);
1854                 count++;
1855         }
1856         unlock_res(res);
1857
1858         RETURN(ldlm_cli_cancel_list_local(cancels, count, cancel_flags));
1859 }
1860 EXPORT_SYMBOL(ldlm_cancel_resource_local);
1861
1862 /**
1863  * Cancel client-side locks from a list and send/prepare cancel RPCs to the
1864  * server.
1865  * If \a req is NULL, send CANCEL request to server with handles of locks
1866  * in the \a cancels. If EARLY_CANCEL is not supported, send CANCEL requests
1867  * separately per lock.
1868  * If \a req is not NULL, put handles of locks in \a cancels into the request
1869  * buffer at the offset \a off.
1870  * Destroy \a cancels at the end.
1871  */
1872 int ldlm_cli_cancel_list(cfs_list_t *cancels, int count,
1873                          struct ptlrpc_request *req, ldlm_cancel_flags_t flags)
1874 {
1875         struct ldlm_lock *lock;
1876         int res = 0;
1877         ENTRY;
1878
1879         if (cfs_list_empty(cancels) || count == 0)
1880                 RETURN(0);
1881
1882         /* XXX: requests (both batched and not) could be sent in parallel.
1883          * Usually it is enough to have just 1 RPC, but it is possible that
1884          * there are too many locks to be cancelled in LRU or on a resource.
1885          * It would also speed up the case when the server does not support
1886          * the feature. */
1887         while (count > 0) {
1888                 LASSERT(!cfs_list_empty(cancels));
1889                 lock = cfs_list_entry(cancels->next, struct ldlm_lock,
1890                                       l_bl_ast);
1891                 LASSERT(lock->l_conn_export);
1892
1893                 if (exp_connect_cancelset(lock->l_conn_export)) {
1894                         res = count;
1895                         if (req)
1896                                 ldlm_cancel_pack(req, cancels, count);
1897                         else
1898                                 res = ldlm_cli_cancel_req(lock->l_conn_export,
1899                                                           cancels, count,
1900                                                           flags);
1901                 } else {
1902                         res = ldlm_cli_cancel_req(lock->l_conn_export,
1903                                                   cancels, 1, flags);
1904                 }
1905
1906                 if (res < 0) {
1907                         CDEBUG_LIMIT(res == -ESHUTDOWN ? D_DLMTRACE : D_ERROR,
1908                                      "ldlm_cli_cancel_list: %d\n", res);
1909                         res = count;
1910                 }
1911
1912                 count -= res;
1913                 ldlm_lock_list_put(cancels, l_bl_ast, res);
1914         }
1915         LASSERT(count == 0);
1916         RETURN(0);
1917 }
1918 EXPORT_SYMBOL(ldlm_cli_cancel_list);
1919
1920 /**
1921  * Cancel all locks on a resource that have 0 readers/writers.
1922  *
1923  * If flags & LDLM_FL_LOCAL_ONLY, throw the locks away without trying
1924  * to notify the server. */
1925 int ldlm_cli_cancel_unused_resource(struct ldlm_namespace *ns,
1926                                     const struct ldlm_res_id *res_id,
1927                                     ldlm_policy_data_t *policy,
1928                                     ldlm_mode_t mode,
1929                                     ldlm_cancel_flags_t flags,
1930                                     void *opaque)
1931 {
1932         struct ldlm_resource *res;
1933         CFS_LIST_HEAD(cancels);
1934         int count;
1935         int rc;
1936         ENTRY;
1937
1938         res = ldlm_resource_get(ns, NULL, res_id, 0, 0);
1939         if (res == NULL) {
1940                 /* This is not a problem. */
1941                 CDEBUG(D_INFO, "No resource "LPU64"\n", res_id->name[0]);
1942                 RETURN(0);
1943         }
1944
1945         LDLM_RESOURCE_ADDREF(res);
1946         count = ldlm_cancel_resource_local(res, &cancels, policy, mode,
1947                                            0, flags | LCF_BL_AST, opaque);
1948         rc = ldlm_cli_cancel_list(&cancels, count, NULL, flags);
1949         if (rc != ELDLM_OK)
1950                 CERROR("canceling unused lock "DLDLMRES": rc = %d\n",
1951                        PLDLMRES(res), rc);
1952
1953         LDLM_RESOURCE_DELREF(res);
1954         ldlm_resource_putref(res);
1955         RETURN(0);
1956 }
1957 EXPORT_SYMBOL(ldlm_cli_cancel_unused_resource);
1958
1959 struct ldlm_cli_cancel_arg {
1960         int     lc_flags;
1961         void   *lc_opaque;
1962 };
1963
1964 static int ldlm_cli_hash_cancel_unused(cfs_hash_t *hs, cfs_hash_bd_t *bd,
1965                                        cfs_hlist_node_t *hnode, void *arg)
1966 {
1967         struct ldlm_resource           *res = cfs_hash_object(hs, hnode);
1968         struct ldlm_cli_cancel_arg     *lc = arg;
1969
1970         ldlm_cli_cancel_unused_resource(ldlm_res_to_ns(res), &res->lr_name,
1971                                         NULL, LCK_MINMODE, lc->lc_flags,
1972                                         lc->lc_opaque);
1973         /* must return 0 for hash iteration */
1974         return 0;
1975 }
1976
1977 /**
1978  * Cancel all locks on a namespace (or a specific resource, if given)
1979  * that have 0 readers/writers.
1980  *
1981  * If flags & LCF_LOCAL, throw the locks away without trying
1982  * to notify the server. */
1983 int ldlm_cli_cancel_unused(struct ldlm_namespace *ns,
1984                            const struct ldlm_res_id *res_id,
1985                            ldlm_cancel_flags_t flags, void *opaque)
1986 {
1987         struct ldlm_cli_cancel_arg arg = {
1988                 .lc_flags       = flags,
1989                 .lc_opaque      = opaque,
1990         };
1991
1992         ENTRY;
1993
1994         if (ns == NULL)
1995                 RETURN(ELDLM_OK);
1996
1997         if (res_id != NULL) {
1998                 RETURN(ldlm_cli_cancel_unused_resource(ns, res_id, NULL,
1999                                                        LCK_MINMODE, flags,
2000                                                        opaque));
2001         } else {
2002                 cfs_hash_for_each_nolock(ns->ns_rs_hash,
2003                                          ldlm_cli_hash_cancel_unused, &arg);
2004                 RETURN(ELDLM_OK);
2005         }
2006 }
2007 EXPORT_SYMBOL(ldlm_cli_cancel_unused);
2008
2009 /* Lock iterators. */
2010
2011 int ldlm_resource_foreach(struct ldlm_resource *res, ldlm_iterator_t iter,
2012                           void *closure)
2013 {
2014         cfs_list_t *tmp, *next;
2015         struct ldlm_lock *lock;
2016         int rc = LDLM_ITER_CONTINUE;
2017
2018         ENTRY;
2019
2020         if (!res)
2021                 RETURN(LDLM_ITER_CONTINUE);
2022
2023         lock_res(res);
2024         cfs_list_for_each_safe(tmp, next, &res->lr_granted) {
2025                 lock = cfs_list_entry(tmp, struct ldlm_lock, l_res_link);
2026
2027                 if (iter(lock, closure) == LDLM_ITER_STOP)
2028                         GOTO(out, rc = LDLM_ITER_STOP);
2029         }
2030
2031         cfs_list_for_each_safe(tmp, next, &res->lr_converting) {
2032                 lock = cfs_list_entry(tmp, struct ldlm_lock, l_res_link);
2033
2034                 if (iter(lock, closure) == LDLM_ITER_STOP)
2035                         GOTO(out, rc = LDLM_ITER_STOP);
2036         }
2037
2038         cfs_list_for_each_safe(tmp, next, &res->lr_waiting) {
2039                 lock = cfs_list_entry(tmp, struct ldlm_lock, l_res_link);
2040
2041                 if (iter(lock, closure) == LDLM_ITER_STOP)
2042                         GOTO(out, rc = LDLM_ITER_STOP);
2043         }
2044  out:
2045         unlock_res(res);
2046         RETURN(rc);
2047 }
2048 EXPORT_SYMBOL(ldlm_resource_foreach);
2049
2050 struct iter_helper_data {
2051         ldlm_iterator_t iter;
2052         void *closure;
2053 };
2054
2055 static int ldlm_iter_helper(struct ldlm_lock *lock, void *closure)
2056 {
2057         struct iter_helper_data *helper = closure;
2058         return helper->iter(lock, helper->closure);
2059 }
2060
2061 static int ldlm_res_iter_helper(cfs_hash_t *hs, cfs_hash_bd_t *bd,
2062                                 cfs_hlist_node_t *hnode, void *arg)
2063
2064 {
2065         struct ldlm_resource *res = cfs_hash_object(hs, hnode);
2066
2067         return ldlm_resource_foreach(res, ldlm_iter_helper, arg) ==
2068                LDLM_ITER_STOP;
2069 }
2070
2071 void ldlm_namespace_foreach(struct ldlm_namespace *ns,
2072                             ldlm_iterator_t iter, void *closure)
2073
2074 {
2075         struct iter_helper_data helper = { iter: iter, closure: closure };
2076
2077         cfs_hash_for_each_nolock(ns->ns_rs_hash,
2078                                  ldlm_res_iter_helper, &helper);
2079
2080 }
2081 EXPORT_SYMBOL(ldlm_namespace_foreach);
2082
2083 /* non-blocking function to manipulate a lock whose cb_data is being put away.
2084  * return  0:  find no resource
2085  *       > 0:  must be LDLM_ITER_STOP/LDLM_ITER_CONTINUE.
2086  *       < 0:  errors
2087  */
2088 int ldlm_resource_iterate(struct ldlm_namespace *ns,
2089                           const struct ldlm_res_id *res_id,
2090                           ldlm_iterator_t iter, void *data)
2091 {
2092         struct ldlm_resource *res;
2093         int rc;
2094         ENTRY;
2095
2096         if (ns == NULL) {
2097                 CERROR("must pass in namespace\n");
2098                 LBUG();
2099         }
2100
2101         res = ldlm_resource_get(ns, NULL, res_id, 0, 0);
2102         if (res == NULL)
2103                 RETURN(0);
2104
2105         LDLM_RESOURCE_ADDREF(res);
2106         rc = ldlm_resource_foreach(res, iter, data);
2107         LDLM_RESOURCE_DELREF(res);
2108         ldlm_resource_putref(res);
2109         RETURN(rc);
2110 }
2111 EXPORT_SYMBOL(ldlm_resource_iterate);
2112
2113 /* Lock replay */
2114
2115 static int ldlm_chain_lock_for_replay(struct ldlm_lock *lock, void *closure)
2116 {
2117         cfs_list_t *list = closure;
2118
2119         /* we use l_pending_chain here, because it's unused on clients. */
2120         LASSERTF(cfs_list_empty(&lock->l_pending_chain),
2121                  "lock %p next %p prev %p\n",
2122                  lock, &lock->l_pending_chain.next,&lock->l_pending_chain.prev);
2123         /* bug 9573: don't replay locks left after eviction, or
2124          * bug 17614: locks being actively cancelled. Get a reference
2125          * on a lock so that it does not disapear under us (e.g. due to cancel)
2126          */
2127         if (!(lock->l_flags & (LDLM_FL_FAILED|LDLM_FL_CANCELING))) {
2128                 cfs_list_add(&lock->l_pending_chain, list);
2129                 LDLM_LOCK_GET(lock);
2130         }
2131
2132         return LDLM_ITER_CONTINUE;
2133 }
2134
2135 static int replay_lock_interpret(const struct lu_env *env,
2136                                  struct ptlrpc_request *req,
2137                                  struct ldlm_async_args *aa, int rc)
2138 {
2139         struct ldlm_lock     *lock;
2140         struct ldlm_reply    *reply;
2141         struct obd_export    *exp;
2142
2143         ENTRY;
2144         atomic_dec(&req->rq_import->imp_replay_inflight);
2145         if (rc != ELDLM_OK)
2146                 GOTO(out, rc);
2147
2148         reply = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
2149         if (reply == NULL)
2150                 GOTO(out, rc = -EPROTO);
2151
2152         lock = ldlm_handle2lock(&aa->lock_handle);
2153         if (!lock) {
2154                 CERROR("received replay ack for unknown local cookie "LPX64
2155                        " remote cookie "LPX64 " from server %s id %s\n",
2156                        aa->lock_handle.cookie, reply->lock_handle.cookie,
2157                        req->rq_export->exp_client_uuid.uuid,
2158                        libcfs_id2str(req->rq_peer));
2159                 GOTO(out, rc = -ESTALE);
2160         }
2161
2162         /* Key change rehash lock in per-export hash with new key */
2163         exp = req->rq_export;
2164         if (exp && exp->exp_lock_hash) {
2165                 /* In the function below, .hs_keycmp resolves to
2166                  * ldlm_export_lock_keycmp() */
2167                 /* coverity[overrun-buffer-val] */
2168                 cfs_hash_rehash_key(exp->exp_lock_hash,
2169                                     &lock->l_remote_handle,
2170                                     &reply->lock_handle,
2171                                     &lock->l_exp_hash);
2172         } else {
2173                 lock->l_remote_handle = reply->lock_handle;
2174         }
2175
2176         LDLM_DEBUG(lock, "replayed lock:");
2177         ptlrpc_import_recovery_state_machine(req->rq_import);
2178         LDLM_LOCK_PUT(lock);
2179 out:
2180         if (rc != ELDLM_OK)
2181                 ptlrpc_connect_import(req->rq_import);
2182
2183         RETURN(rc);
2184 }
2185
2186 static int replay_one_lock(struct obd_import *imp, struct ldlm_lock *lock)
2187 {
2188         struct ptlrpc_request *req;
2189         struct ldlm_async_args *aa;
2190         struct ldlm_request   *body;
2191         int flags;
2192         ENTRY;
2193
2194
2195         /* Bug 11974: Do not replay a lock which is actively being canceled */
2196         if (ldlm_is_canceling(lock)) {
2197                 LDLM_DEBUG(lock, "Not replaying canceled lock:");
2198                 RETURN(0);
2199         }
2200
2201         /* If this is reply-less callback lock, we cannot replay it, since
2202          * server might have long dropped it, but notification of that event was
2203          * lost by network. (and server granted conflicting lock already) */
2204         if (ldlm_is_cancel_on_block(lock)) {
2205                 LDLM_DEBUG(lock, "Not replaying reply-less lock:");
2206                 ldlm_lock_cancel(lock);
2207                 RETURN(0);
2208         }
2209
2210         /*
2211          * If granted mode matches the requested mode, this lock is granted.
2212          *
2213          * If they differ, but we have a granted mode, then we were granted
2214          * one mode and now want another: ergo, converting.
2215          *
2216          * If we haven't been granted anything and are on a resource list,
2217          * then we're blocked/waiting.
2218          *
2219          * If we haven't been granted anything and we're NOT on a resource list,
2220          * then we haven't got a reply yet and don't have a known disposition.
2221          * This happens whenever a lock enqueue is the request that triggers
2222          * recovery.
2223          */
2224         if (lock->l_granted_mode == lock->l_req_mode)
2225                 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_GRANTED;
2226         else if (lock->l_granted_mode)
2227                 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_CONV;
2228         else if (!cfs_list_empty(&lock->l_res_link))
2229                 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_WAIT;
2230         else
2231                 flags = LDLM_FL_REPLAY;
2232
2233         req = ptlrpc_request_alloc_pack(imp, &RQF_LDLM_ENQUEUE,
2234                                         LUSTRE_DLM_VERSION, LDLM_ENQUEUE);
2235         if (req == NULL)
2236                 RETURN(-ENOMEM);
2237
2238         /* We're part of recovery, so don't wait for it. */
2239         req->rq_send_state = LUSTRE_IMP_REPLAY_LOCKS;
2240
2241         body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2242         ldlm_lock2desc(lock, &body->lock_desc);
2243         body->lock_flags = ldlm_flags_to_wire(flags);
2244
2245         ldlm_lock2handle(lock, &body->lock_handle[0]);
2246         if (lock->l_lvb_len > 0)
2247                 req_capsule_extend(&req->rq_pill, &RQF_LDLM_ENQUEUE_LVB);
2248         req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
2249                              lock->l_lvb_len);
2250         ptlrpc_request_set_replen(req);
2251         /* notify the server we've replayed all requests.
2252          * also, we mark the request to be put on a dedicated
2253          * queue to be processed after all request replayes.
2254          * bug 6063 */
2255         lustre_msg_set_flags(req->rq_reqmsg, MSG_REQ_REPLAY_DONE);
2256
2257         LDLM_DEBUG(lock, "replaying lock:");
2258
2259         atomic_inc(&req->rq_import->imp_replay_inflight);
2260         CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
2261         aa = ptlrpc_req_async_args(req);
2262         aa->lock_handle = body->lock_handle[0];
2263         req->rq_interpret_reply = (ptlrpc_interpterer_t)replay_lock_interpret;
2264         ptlrpcd_add_req(req, PDL_POLICY_LOCAL, -1);
2265
2266         RETURN(0);
2267 }
2268
2269 /**
2270  * Cancel as many unused locks as possible before replay. since we are
2271  * in recovery, we can't wait for any outstanding RPCs to send any RPC
2272  * to the server.
2273  *
2274  * Called only in recovery before replaying locks. there is no need to
2275  * replay locks that are unused. since the clients may hold thousands of
2276  * cached unused locks, dropping the unused locks can greatly reduce the
2277  * load on the servers at recovery time.
2278  */
2279 static void ldlm_cancel_unused_locks_for_replay(struct ldlm_namespace *ns)
2280 {
2281         int canceled;
2282         CFS_LIST_HEAD(cancels);
2283
2284         CDEBUG(D_DLMTRACE, "Dropping as many unused locks as possible before"
2285                            "replay for namespace %s (%d)\n",
2286                            ldlm_ns_name(ns), ns->ns_nr_unused);
2287
2288         /* We don't need to care whether or not LRU resize is enabled
2289          * because the LDLM_CANCEL_NO_WAIT policy doesn't use the
2290          * count parameter */
2291         canceled = ldlm_cancel_lru_local(ns, &cancels, ns->ns_nr_unused, 0,
2292                                          LCF_LOCAL, LDLM_CANCEL_NO_WAIT);
2293
2294         CDEBUG(D_DLMTRACE, "Canceled %d unused locks from namespace %s\n",
2295                            canceled, ldlm_ns_name(ns));
2296 }
2297
2298 int ldlm_replay_locks(struct obd_import *imp)
2299 {
2300         struct ldlm_namespace *ns = imp->imp_obd->obd_namespace;
2301         CFS_LIST_HEAD(list);
2302         struct ldlm_lock *lock, *next;
2303         int rc = 0;
2304
2305         ENTRY;
2306
2307         LASSERT(atomic_read(&imp->imp_replay_inflight) == 0);
2308
2309         /* don't replay locks if import failed recovery */
2310         if (imp->imp_vbr_failed)
2311                 RETURN(0);
2312
2313         /* ensure this doesn't fall to 0 before all have been queued */
2314         atomic_inc(&imp->imp_replay_inflight);
2315
2316         if (ldlm_cancel_unused_locks_before_replay)
2317                 ldlm_cancel_unused_locks_for_replay(ns);
2318
2319         ldlm_namespace_foreach(ns, ldlm_chain_lock_for_replay, &list);
2320
2321         cfs_list_for_each_entry_safe(lock, next, &list, l_pending_chain) {
2322                 cfs_list_del_init(&lock->l_pending_chain);
2323                 if (rc) {
2324                         LDLM_LOCK_RELEASE(lock);
2325                         continue; /* or try to do the rest? */
2326                 }
2327                 rc = replay_one_lock(imp, lock);
2328                 LDLM_LOCK_RELEASE(lock);
2329         }
2330
2331         atomic_dec(&imp->imp_replay_inflight);
2332
2333         RETURN(rc);
2334 }
2335 EXPORT_SYMBOL(ldlm_replay_locks);