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