Whamcloud - gitweb
LU-2901 ldlm: fix resource/fid check, use DLDLMRES
[fs/lustre-release.git] / lustre / ldlm / ldlm_request.c
1 /*
2  * GPL HEADER START
3  *
4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
5  *
6  * This program is free software; you can redistribute it and/or modify
7  * it under the terms of the GNU General Public License version 2 only,
8  * as published by the Free Software Foundation.
9  *
10  * This program is distributed in the hope that it will be useful, but
11  * WITHOUT ANY WARRANTY; without even the implied warranty of
12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
13  * General Public License version 2 for more details (a copy is included
14  * in the LICENSE file that accompanied this code).
15  *
16  * You should have received a copy of the GNU General Public License
17  * version 2 along with this program; If not, see
18  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
19  *
20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21  * CA 95054 USA or visit www.sun.com if you need additional information or
22  * have any questions.
23  *
24  * GPL HEADER END
25  */
26 /*
27  * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
28  * Use is subject to license terms.
29  *
30  * Copyright (c) 2010, 2013, Intel Corporation.
31  */
32 /*
33  * This file is part of Lustre, http://www.lustre.org/
34  * Lustre is a trademark of Sun Microsystems, Inc.
35  */
36 /**
37  * This file contains Asynchronous System Trap (AST) handlers and related
38  * LDLM request-processing routines.
39  *
40  * An AST is a callback issued on a lock when its state is changed. There are
41  * several different types of ASTs (callbacks) registered for each lock:
42  *
43  * - completion AST: when a lock is enqueued by some process, but cannot be
44  *   granted immediately due to other conflicting locks on the same resource,
45  *   the completion AST is sent to notify the caller when the lock is
46  *   eventually granted
47  *
48  * - blocking AST: when a lock is granted to some process, if another process
49  *   enqueues a conflicting (blocking) lock on a resource, a blocking AST is
50  *   sent to notify the holder(s) of the lock(s) of the conflicting lock
51  *   request. The lock holder(s) must release their lock(s) on that resource in
52  *   a timely manner or be evicted by the server.
53  *
54  * - glimpse AST: this is used when a process wants information about a lock
55  *   (i.e. the lock value block (LVB)) but does not necessarily require holding
56  *   the lock. If the resource is locked, the lock holder(s) are sent glimpse
57  *   ASTs and the LVB is returned to the caller, and lock holder(s) may CANCEL
58  *   their lock(s) if they are idle. If the resource is not locked, the server
59  *   may grant the lock.
60  */
61
62 #define DEBUG_SUBSYSTEM S_LDLM
63 #ifndef __KERNEL__
64 #include <signal.h>
65 #include <liblustre.h>
66 #endif
67
68 #include <lustre_dlm.h>
69 #include <obd_class.h>
70 #include <obd.h>
71
72 #include "ldlm_internal.h"
73
74 int ldlm_enqueue_min = OBD_TIMEOUT_DEFAULT;
75 CFS_MODULE_PARM(ldlm_enqueue_min, "i", int, 0644,
76                 "lock enqueue timeout minimum");
77
78 /* in client side, whether the cached locks will be canceled before replay */
79 unsigned int ldlm_cancel_unused_locks_before_replay = 1;
80
81 static void interrupted_completion_wait(void *data)
82 {
83 }
84
85 struct lock_wait_data {
86         struct ldlm_lock *lwd_lock;
87         __u32             lwd_conn_cnt;
88 };
89
90 struct ldlm_async_args {
91         struct lustre_handle lock_handle;
92 };
93
94 int ldlm_expired_completion_wait(void *data)
95 {
96         struct lock_wait_data *lwd = data;
97         struct ldlm_lock *lock = lwd->lwd_lock;
98         struct obd_import *imp;
99         struct obd_device *obd;
100
101         ENTRY;
102         if (lock->l_conn_export == NULL) {
103                 static cfs_time_t next_dump = 0, last_dump = 0;
104
105                 if (ptlrpc_check_suspend())
106                         RETURN(0);
107
108                 LCONSOLE_WARN("lock timed out (enqueued at "CFS_TIME_T", "
109                               CFS_DURATION_T"s ago)\n",
110                               lock->l_last_activity,
111                               cfs_time_sub(cfs_time_current_sec(),
112                                            lock->l_last_activity));
113                 LDLM_DEBUG(lock, "lock timed out (enqueued at "CFS_TIME_T", "
114                            CFS_DURATION_T"s ago); not entering recovery in "
115                            "server code, just going back to sleep",
116                            lock->l_last_activity,
117                            cfs_time_sub(cfs_time_current_sec(),
118                                         lock->l_last_activity));
119                 if (cfs_time_after(cfs_time_current(), next_dump)) {
120                         last_dump = next_dump;
121                         next_dump = cfs_time_shift(300);
122                         ldlm_namespace_dump(D_DLMTRACE,
123                                             ldlm_lock_to_ns(lock));
124                         if (last_dump == 0)
125                                 libcfs_debug_dumplog();
126                 }
127                 RETURN(0);
128         }
129
130         obd = lock->l_conn_export->exp_obd;
131         imp = obd->u.cli.cl_import;
132         ptlrpc_fail_import(imp, lwd->lwd_conn_cnt);
133         LDLM_ERROR(lock, "lock timed out (enqueued at "CFS_TIME_T", "
134                   CFS_DURATION_T"s ago), entering recovery for %s@%s",
135                   lock->l_last_activity,
136                   cfs_time_sub(cfs_time_current_sec(), lock->l_last_activity),
137                   obd2cli_tgt(obd), imp->imp_connection->c_remote_uuid.uuid);
138
139         RETURN(0);
140 }
141 EXPORT_SYMBOL(ldlm_expired_completion_wait);
142
143 /* We use the same basis for both server side and client side functions
144    from a single node. */
145 int ldlm_get_enq_timeout(struct ldlm_lock *lock)
146 {
147         int timeout = at_get(ldlm_lock_to_ns_at(lock));
148         if (AT_OFF)
149                 return obd_timeout / 2;
150         /* Since these are non-updating timeouts, we should be conservative.
151            It would be nice to have some kind of "early reply" mechanism for
152            lock callbacks too... */
153         timeout = min_t(int, at_max, timeout + (timeout >> 1)); /* 150% */
154         return max(timeout, ldlm_enqueue_min);
155 }
156 EXPORT_SYMBOL(ldlm_get_enq_timeout);
157
158 /**
159  * Helper function for ldlm_completion_ast(), updating timings when lock is
160  * actually granted.
161  */
162 static int ldlm_completion_tail(struct ldlm_lock *lock)
163 {
164         long delay;
165         int  result;
166
167         if (lock->l_flags & (LDLM_FL_DESTROYED | LDLM_FL_FAILED)) {
168                 LDLM_DEBUG(lock, "client-side enqueue: destroyed");
169                 result = -EIO;
170         } else {
171                 delay = cfs_time_sub(cfs_time_current_sec(),
172                                      lock->l_last_activity);
173                 LDLM_DEBUG(lock, "client-side enqueue: granted after "
174                            CFS_DURATION_T"s", delay);
175
176                 /* Update our time estimate */
177                 at_measured(ldlm_lock_to_ns_at(lock),
178                             delay);
179                 result = 0;
180         }
181         return result;
182 }
183
184 /**
185  * Implementation of ->l_completion_ast() for a client, that doesn't wait
186  * until lock is granted. Suitable for locks enqueued through ptlrpcd, of
187  * other threads that cannot block for long.
188  */
189 int ldlm_completion_ast_async(struct ldlm_lock *lock, __u64 flags, void *data)
190 {
191         ENTRY;
192
193         if (flags == LDLM_FL_WAIT_NOREPROC) {
194                 LDLM_DEBUG(lock, "client-side enqueue waiting on pending lock");
195                 RETURN(0);
196         }
197
198         if (!(flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED |
199                        LDLM_FL_BLOCK_CONV))) {
200                 cfs_waitq_signal(&lock->l_waitq);
201                 RETURN(ldlm_completion_tail(lock));
202         }
203
204         LDLM_DEBUG(lock, "client-side enqueue returned a blocked lock, "
205                    "going forward");
206         ldlm_reprocess_all(lock->l_resource);
207         RETURN(0);
208 }
209 EXPORT_SYMBOL(ldlm_completion_ast_async);
210
211 /**
212  * Generic LDLM "completion" AST. This is called in several cases:
213  *
214  *     - when a reply to an ENQUEUE RPC is received from the server
215  *       (ldlm_cli_enqueue_fini()). Lock might be granted or not granted at
216  *       this point (determined by flags);
217  *
218  *     - when LDLM_CP_CALLBACK RPC comes to client to notify it that lock has
219  *       been granted;
220  *
221  *     - when ldlm_lock_match(LDLM_FL_LVB_READY) is about to wait until lock
222  *       gets correct lvb;
223  *
224  *     - to force all locks when resource is destroyed (cleanup_resource());
225  *
226  *     - during lock conversion (not used currently).
227  *
228  * If lock is not granted in the first case, this function waits until second
229  * or penultimate cases happen in some other thread.
230  *
231  */
232 int ldlm_completion_ast(struct ldlm_lock *lock, __u64 flags, void *data)
233 {
234         /* XXX ALLOCATE - 160 bytes */
235         struct lock_wait_data lwd;
236         struct obd_device *obd;
237         struct obd_import *imp = NULL;
238         struct l_wait_info lwi;
239         __u32 timeout;
240         int rc = 0;
241         ENTRY;
242
243         if (flags == LDLM_FL_WAIT_NOREPROC) {
244                 LDLM_DEBUG(lock, "client-side enqueue waiting on pending lock");
245                 goto noreproc;
246         }
247
248         if (!(flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED |
249                        LDLM_FL_BLOCK_CONV))) {
250                 cfs_waitq_signal(&lock->l_waitq);
251                 RETURN(0);
252         }
253
254         LDLM_DEBUG(lock, "client-side enqueue returned a blocked lock, "
255                    "sleeping");
256
257 noreproc:
258
259         obd = class_exp2obd(lock->l_conn_export);
260
261         /* if this is a local lock, then there is no import */
262         if (obd != NULL) {
263                 imp = obd->u.cli.cl_import;
264         }
265
266         /* Wait a long time for enqueue - server may have to callback a
267            lock from another client.  Server will evict the other client if it
268            doesn't respond reasonably, and then give us the lock. */
269         timeout = ldlm_get_enq_timeout(lock) * 2;
270
271         lwd.lwd_lock = lock;
272
273         if (lock->l_flags & LDLM_FL_NO_TIMEOUT) {
274                 LDLM_DEBUG(lock, "waiting indefinitely because of NO_TIMEOUT");
275                 lwi = LWI_INTR(interrupted_completion_wait, &lwd);
276         } else {
277                 lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(timeout),
278                                        ldlm_expired_completion_wait,
279                                        interrupted_completion_wait, &lwd);
280         }
281
282         if (imp != NULL) {
283                 spin_lock(&imp->imp_lock);
284                 lwd.lwd_conn_cnt = imp->imp_conn_cnt;
285                 spin_unlock(&imp->imp_lock);
286         }
287
288         if (ns_is_client(ldlm_lock_to_ns(lock)) &&
289             OBD_FAIL_CHECK_RESET(OBD_FAIL_LDLM_INTR_CP_AST,
290                                  OBD_FAIL_LDLM_CP_BL_RACE | OBD_FAIL_ONCE)) {
291                 lock->l_flags |= LDLM_FL_FAIL_LOC;
292                 rc = -EINTR;
293         } else {
294                 /* Go to sleep until the lock is granted or cancelled. */
295                 rc = l_wait_event(lock->l_waitq,
296                                   is_granted_or_cancelled(lock), &lwi);
297         }
298
299         if (rc) {
300                 LDLM_DEBUG(lock, "client-side enqueue waking up: failed (%d)",
301                            rc);
302                 RETURN(rc);
303         }
304
305         RETURN(ldlm_completion_tail(lock));
306 }
307 EXPORT_SYMBOL(ldlm_completion_ast);
308
309 /**
310  * A helper to build a blocking AST function
311  *
312  * Perform a common operation for blocking ASTs:
313  * defferred lock cancellation.
314  *
315  * \param lock the lock blocking or canceling AST was called on
316  * \retval 0
317  * \see mdt_blocking_ast
318  * \see ldlm_blocking_ast
319  */
320 int ldlm_blocking_ast_nocheck(struct ldlm_lock *lock)
321 {
322         int do_ast;
323         ENTRY;
324
325         lock->l_flags |= LDLM_FL_CBPENDING;
326         do_ast = (!lock->l_readers && !lock->l_writers);
327         unlock_res_and_lock(lock);
328
329         if (do_ast) {
330                 struct lustre_handle lockh;
331                 int rc;
332
333                 LDLM_DEBUG(lock, "already unused, calling ldlm_cli_cancel");
334                 ldlm_lock2handle(lock, &lockh);
335                 rc = ldlm_cli_cancel(&lockh, LCF_ASYNC);
336                 if (rc < 0)
337                         CERROR("ldlm_cli_cancel: %d\n", rc);
338         } else {
339                 LDLM_DEBUG(lock, "Lock still has references, will be "
340                            "cancelled later");
341         }
342         RETURN(0);
343 }
344 EXPORT_SYMBOL(ldlm_blocking_ast_nocheck);
345
346 /**
347  * Server blocking AST
348  *
349  * ->l_blocking_ast() callback for LDLM locks acquired by server-side
350  * OBDs.
351  *
352  * \param lock the lock which blocks a request or cancelling lock
353  * \param desc unused
354  * \param data unused
355  * \param flag indicates whether this cancelling or blocking callback
356  * \retval 0
357  * \see ldlm_blocking_ast_nocheck
358  */
359 int ldlm_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
360                       void *data, int flag)
361 {
362         ENTRY;
363
364         if (flag == LDLM_CB_CANCELING) {
365                 /* Don't need to do anything here. */
366                 RETURN(0);
367         }
368
369         lock_res_and_lock(lock);
370         /* Get this: if ldlm_blocking_ast is racing with intent_policy, such
371          * that ldlm_blocking_ast is called just before intent_policy method
372          * takes the lr_lock, then by the time we get the lock, we might not
373          * be the correct blocking function anymore.  So check, and return
374          * early, if so. */
375         if (lock->l_blocking_ast != ldlm_blocking_ast) {
376                 unlock_res_and_lock(lock);
377                 RETURN(0);
378         }
379         RETURN(ldlm_blocking_ast_nocheck(lock));
380 }
381 EXPORT_SYMBOL(ldlm_blocking_ast);
382
383 /**
384  * ->l_glimpse_ast() for DLM extent locks acquired on the server-side. See
385  * comment in filter_intent_policy() on why you may need this.
386  */
387 int ldlm_glimpse_ast(struct ldlm_lock *lock, void *reqp)
388 {
389         /*
390          * Returning -ELDLM_NO_LOCK_DATA actually works, but the reason for
391          * that is rather subtle: with OST-side locking, it may so happen that
392          * _all_ extent locks are held by the OST. If client wants to obtain
393          * current file size it calls ll{,u}_glimpse_size(), and (as locks are
394          * on the server), dummy glimpse callback fires and does
395          * nothing. Client still receives correct file size due to the
396          * following fragment in filter_intent_policy():
397          *
398          * rc = l->l_glimpse_ast(l, NULL); // this will update the LVB
399          * if (rc != 0 && res->lr_namespace->ns_lvbo &&
400          *     res->lr_namespace->ns_lvbo->lvbo_update) {
401          *         res->lr_namespace->ns_lvbo->lvbo_update(res, NULL, 0, 1);
402          * }
403          *
404          * that is, after glimpse_ast() fails, filter_lvbo_update() runs, and
405          * returns correct file size to the client.
406          */
407         return -ELDLM_NO_LOCK_DATA;
408 }
409 EXPORT_SYMBOL(ldlm_glimpse_ast);
410
411 /**
412  * Enqueue a local lock (typically on a server).
413  */
414 int ldlm_cli_enqueue_local(struct ldlm_namespace *ns,
415                            const struct ldlm_res_id *res_id,
416                            ldlm_type_t type, ldlm_policy_data_t *policy,
417                            ldlm_mode_t mode, __u64 *flags,
418                            ldlm_blocking_callback blocking,
419                            ldlm_completion_callback completion,
420                            ldlm_glimpse_callback glimpse,
421                            void *data, __u32 lvb_len, enum lvb_type lvb_type,
422                            const __u64 *client_cookie,
423                            struct lustre_handle *lockh)
424 {
425         struct ldlm_lock *lock;
426         int err;
427         const struct ldlm_callback_suite cbs = { .lcs_completion = completion,
428                                                  .lcs_blocking   = blocking,
429                                                  .lcs_glimpse    = glimpse,
430         };
431         ENTRY;
432
433         LASSERT(!(*flags & LDLM_FL_REPLAY));
434         if (unlikely(ns_is_client(ns))) {
435                 CERROR("Trying to enqueue local lock in a shadow namespace\n");
436                 LBUG();
437         }
438
439         lock = ldlm_lock_create(ns, res_id, type, mode, &cbs, data, lvb_len,
440                                 lvb_type);
441         if (unlikely(!lock))
442                 GOTO(out_nolock, err = -ENOMEM);
443
444         ldlm_lock2handle(lock, lockh);
445
446         /* NB: we don't have any lock now (lock_res_and_lock)
447          * because it's a new lock */
448         ldlm_lock_addref_internal_nolock(lock, mode);
449         lock->l_flags |= LDLM_FL_LOCAL;
450         if (*flags & LDLM_FL_ATOMIC_CB)
451                 lock->l_flags |= LDLM_FL_ATOMIC_CB;
452
453         if (policy != NULL)
454                 lock->l_policy_data = *policy;
455         if (client_cookie != NULL)
456                 lock->l_client_cookie = *client_cookie;
457         if (type == LDLM_EXTENT)
458                 lock->l_req_extent = policy->l_extent;
459
460         err = ldlm_lock_enqueue(ns, &lock, policy, flags);
461         if (unlikely(err != ELDLM_OK))
462                 GOTO(out, err);
463
464         if (policy != NULL)
465                 *policy = lock->l_policy_data;
466
467         if (lock->l_completion_ast)
468                 lock->l_completion_ast(lock, *flags, NULL);
469
470         LDLM_DEBUG(lock, "client-side local enqueue handler, new lock created");
471         EXIT;
472  out:
473         LDLM_LOCK_RELEASE(lock);
474  out_nolock:
475         return err;
476 }
477 EXPORT_SYMBOL(ldlm_cli_enqueue_local);
478
479 static void failed_lock_cleanup(struct ldlm_namespace *ns,
480                                 struct ldlm_lock *lock, int mode)
481 {
482         int need_cancel = 0;
483
484         /* Set a flag to prevent us from sending a CANCEL (bug 407) */
485         lock_res_and_lock(lock);
486         /* Check that lock is not granted or failed, we might race. */
487         if ((lock->l_req_mode != lock->l_granted_mode) &&
488             !(lock->l_flags & LDLM_FL_FAILED)) {
489                 /* Make sure that this lock will not be found by raced
490                  * bl_ast and -EINVAL reply is sent to server anyways.
491                  * bug 17645 */
492                 lock->l_flags |= LDLM_FL_LOCAL_ONLY | LDLM_FL_FAILED |
493                                  LDLM_FL_ATOMIC_CB | LDLM_FL_CBPENDING;
494                 need_cancel = 1;
495         }
496         unlock_res_and_lock(lock);
497
498         if (need_cancel)
499                 LDLM_DEBUG(lock,
500                            "setting FL_LOCAL_ONLY | LDLM_FL_FAILED | "
501                            "LDLM_FL_ATOMIC_CB | LDLM_FL_CBPENDING");
502         else
503                 LDLM_DEBUG(lock, "lock was granted or failed in race");
504
505         ldlm_lock_decref_internal(lock, mode);
506
507         /* XXX - HACK because we shouldn't call ldlm_lock_destroy()
508          *       from llite/file.c/ll_file_flock(). */
509         /* This code makes for the fact that we do not have blocking handler on
510          * a client for flock locks. As such this is the place where we must
511          * completely kill failed locks. (interrupted and those that
512          * were waiting to be granted when server evicted us. */
513         if (lock->l_resource->lr_type == LDLM_FLOCK) {
514                 lock_res_and_lock(lock);
515                 ldlm_resource_unlink_lock(lock);
516                 ldlm_lock_destroy_nolock(lock);
517                 unlock_res_and_lock(lock);
518         }
519 }
520
521 /**
522  * Finishing portion of client lock enqueue code.
523  *
524  * Called after receiving reply from server.
525  */
526 int ldlm_cli_enqueue_fini(struct obd_export *exp, struct ptlrpc_request *req,
527                           ldlm_type_t type, __u8 with_policy, ldlm_mode_t mode,
528                           __u64 *flags, void *lvb, __u32 lvb_len,
529                           struct lustre_handle *lockh,int rc)
530 {
531         struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
532         int is_replay = *flags & LDLM_FL_REPLAY;
533         struct ldlm_lock *lock;
534         struct ldlm_reply *reply;
535         int cleanup_phase = 1;
536         int size = 0;
537         ENTRY;
538
539         lock = ldlm_handle2lock(lockh);
540         /* ldlm_cli_enqueue is holding a reference on this lock. */
541         if (!lock) {
542                 LASSERT(type == LDLM_FLOCK);
543                 RETURN(-ENOLCK);
544         }
545
546         LASSERTF(ergo(lvb_len != 0, lvb_len == lock->l_lvb_len),
547                  "lvb_len = %d, l_lvb_len = %d\n", lvb_len, lock->l_lvb_len);
548
549         if (rc != ELDLM_OK) {
550                 LASSERT(!is_replay);
551                 LDLM_DEBUG(lock, "client-side enqueue END (%s)",
552                            rc == ELDLM_LOCK_ABORTED ? "ABORTED" : "FAILED");
553
554                 if (rc != ELDLM_LOCK_ABORTED)
555                         GOTO(cleanup, rc);
556         }
557
558         /* Before we return, swab the reply */
559         reply = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
560         if (reply == NULL)
561                 GOTO(cleanup, rc = -EPROTO);
562
563         if (lvb_len != 0) {
564                 LASSERT(lvb != NULL);
565
566                 size = req_capsule_get_size(&req->rq_pill, &RMF_DLM_LVB,
567                                             RCL_SERVER);
568                 if (size < 0) {
569                         LDLM_ERROR(lock, "Fail to get lvb_len, rc = %d", size);
570                         GOTO(cleanup, rc = size);
571                 } else if (unlikely(size > lvb_len)) {
572                         LDLM_ERROR(lock, "Replied LVB is larger than "
573                                    "expectation, expected = %d, replied = %d",
574                                    lvb_len, size);
575                         GOTO(cleanup, rc = -EINVAL);
576                 }
577         }
578
579         if (rc == ELDLM_LOCK_ABORTED) {
580                 if (lvb_len != 0)
581                         rc = ldlm_fill_lvb(lock, &req->rq_pill, RCL_SERVER,
582                                            lvb, size);
583                 GOTO(cleanup, rc = (rc != 0 ? rc : ELDLM_LOCK_ABORTED));
584         }
585
586         /* lock enqueued on the server */
587         cleanup_phase = 0;
588
589         lock_res_and_lock(lock);
590         /* Key change rehash lock in per-export hash with new key */
591         if (exp->exp_lock_hash) {
592                 /* In the function below, .hs_keycmp resolves to
593                  * ldlm_export_lock_keycmp() */
594                 /* coverity[overrun-buffer-val] */
595                 cfs_hash_rehash_key(exp->exp_lock_hash,
596                                     &lock->l_remote_handle,
597                                     &reply->lock_handle,
598                                     &lock->l_exp_hash);
599         } else {
600                 lock->l_remote_handle = reply->lock_handle;
601         }
602
603         *flags = ldlm_flags_from_wire(reply->lock_flags);
604         lock->l_flags |= ldlm_flags_from_wire(reply->lock_flags &
605                                               LDLM_INHERIT_FLAGS);
606         /* move NO_TIMEOUT flag to the lock to force ldlm_lock_match()
607          * to wait with no timeout as well */
608         lock->l_flags |= ldlm_flags_from_wire(reply->lock_flags &
609                                               LDLM_FL_NO_TIMEOUT);
610         unlock_res_and_lock(lock);
611
612         CDEBUG(D_INFO, "local: %p, remote cookie: "LPX64", flags: 0x%llx\n",
613                lock, reply->lock_handle.cookie, *flags);
614
615         /* If enqueue returned a blocked lock but the completion handler has
616          * already run, then it fixed up the resource and we don't need to do it
617          * again. */
618         if ((*flags) & LDLM_FL_LOCK_CHANGED) {
619                 int newmode = reply->lock_desc.l_req_mode;
620                 LASSERT(!is_replay);
621                 if (newmode && newmode != lock->l_req_mode) {
622                         LDLM_DEBUG(lock, "server returned different mode %s",
623                                    ldlm_lockname[newmode]);
624                         lock->l_req_mode = newmode;
625                 }
626
627                 if (!ldlm_res_eq(&reply->lock_desc.l_resource.lr_name,
628                                  &lock->l_resource->lr_name)) {
629                         CDEBUG(D_INFO, "remote intent success, locking "DLDLMRES
630                                        " instead of "DLDLMRES"\n",
631                                PLDLMRES(&reply->lock_desc.l_resource),
632                                PLDLMRES(lock->l_resource));
633
634                         rc = ldlm_lock_change_resource(ns, lock,
635                                         &reply->lock_desc.l_resource.lr_name);
636                         if (rc || lock->l_resource == NULL)
637                                 GOTO(cleanup, rc = -ENOMEM);
638                         LDLM_DEBUG(lock, "client-side enqueue, new resource");
639                 }
640                 if (with_policy)
641                         if (!(type == LDLM_IBITS &&
642                               !(exp_connect_flags(exp) & OBD_CONNECT_IBITS)))
643                                 /* We assume lock type cannot change on server*/
644                                 ldlm_convert_policy_to_local(exp,
645                                                 lock->l_resource->lr_type,
646                                                 &reply->lock_desc.l_policy_data,
647                                                 &lock->l_policy_data);
648                 if (type != LDLM_PLAIN)
649                         LDLM_DEBUG(lock,"client-side enqueue, new policy data");
650         }
651
652         if ((*flags) & LDLM_FL_AST_SENT ||
653             /* Cancel extent locks as soon as possible on a liblustre client,
654              * because it cannot handle asynchronous ASTs robustly (see
655              * bug 7311). */
656             (LIBLUSTRE_CLIENT && type == LDLM_EXTENT)) {
657                 lock_res_and_lock(lock);
658                 lock->l_flags |= LDLM_FL_CBPENDING |  LDLM_FL_BL_AST;
659                 unlock_res_and_lock(lock);
660                 LDLM_DEBUG(lock, "enqueue reply includes blocking AST");
661         }
662
663         /* If the lock has already been granted by a completion AST, don't
664          * clobber the LVB with an older one. */
665         if (lvb_len != 0) {
666                 /* We must lock or a racing completion might update lvb without
667                  * letting us know and we'll clobber the correct value.
668                  * Cannot unlock after the check either, a that still leaves
669                  * a tiny window for completion to get in */
670                 lock_res_and_lock(lock);
671                 if (lock->l_req_mode != lock->l_granted_mode)
672                         rc = ldlm_fill_lvb(lock, &req->rq_pill, RCL_SERVER,
673                                            lock->l_lvb_data, size);
674                 unlock_res_and_lock(lock);
675                 if (rc < 0) {
676                         cleanup_phase = 1;
677                         GOTO(cleanup, rc);
678                 }
679         }
680
681         if (!is_replay) {
682                 rc = ldlm_lock_enqueue(ns, &lock, NULL, flags);
683                 if (lock->l_completion_ast != NULL) {
684                         int err = lock->l_completion_ast(lock, *flags, NULL);
685                         if (!rc)
686                                 rc = err;
687                         if (rc)
688                                 cleanup_phase = 1;
689                 }
690         }
691
692         if (lvb_len && lvb != NULL) {
693                 /* Copy the LVB here, and not earlier, because the completion
694                  * AST (if any) can override what we got in the reply */
695                 memcpy(lvb, lock->l_lvb_data, lvb_len);
696         }
697
698         LDLM_DEBUG(lock, "client-side enqueue END");
699         EXIT;
700 cleanup:
701         if (cleanup_phase == 1 && rc)
702                 failed_lock_cleanup(ns, lock, mode);
703         /* Put lock 2 times, the second reference is held by ldlm_cli_enqueue */
704         LDLM_LOCK_PUT(lock);
705         LDLM_LOCK_RELEASE(lock);
706         return rc;
707 }
708 EXPORT_SYMBOL(ldlm_cli_enqueue_fini);
709
710 /**
711  * Estimate number of lock handles that would fit into request of given
712  * size.  PAGE_SIZE-512 is to allow TCP/IP and LNET headers to fit into
713  * a single page on the send/receive side. XXX: 512 should be changed to
714  * more adequate value.
715  */
716 static inline int ldlm_req_handles_avail(int req_size, int off)
717 {
718         int avail;
719
720         avail = min_t(int, LDLM_MAXREQSIZE, PAGE_CACHE_SIZE - 512) - req_size;
721         if (likely(avail >= 0))
722                 avail /= (int)sizeof(struct lustre_handle);
723         else
724                 avail = 0;
725         avail += LDLM_LOCKREQ_HANDLES - off;
726
727         return avail;
728 }
729
730 static inline int ldlm_capsule_handles_avail(struct req_capsule *pill,
731                                              enum req_location loc,
732                                              int off)
733 {
734         int size = req_capsule_msg_size(pill, loc);
735         return ldlm_req_handles_avail(size, off);
736 }
737
738 static inline int ldlm_format_handles_avail(struct obd_import *imp,
739                                             const struct req_format *fmt,
740                                             enum req_location loc, int off)
741 {
742         int size = req_capsule_fmt_size(imp->imp_msg_magic, fmt, loc);
743         return ldlm_req_handles_avail(size, off);
744 }
745
746 /**
747  * Cancel LRU locks and pack them into the enqueue request. Pack there the given
748  * \a count locks in \a cancels.
749  *
750  * This is to be called by functions preparing their own requests that
751  * might contain lists of locks to cancel in addition to actual operation
752  * that needs to be performed.
753  */
754 int ldlm_prep_elc_req(struct obd_export *exp, struct ptlrpc_request *req,
755                       int version, int opc, int canceloff,
756                       cfs_list_t *cancels, int count)
757 {
758         struct ldlm_namespace   *ns = exp->exp_obd->obd_namespace;
759         struct req_capsule      *pill = &req->rq_pill;
760         struct ldlm_request     *dlm = NULL;
761         int flags, avail, to_free, pack = 0;
762         CFS_LIST_HEAD(head);
763         int rc;
764         ENTRY;
765
766         if (cancels == NULL)
767                 cancels = &head;
768         if (ns_connect_cancelset(ns)) {
769                 /* Estimate the amount of available space in the request. */
770                 req_capsule_filled_sizes(pill, RCL_CLIENT);
771                 avail = ldlm_capsule_handles_avail(pill, RCL_CLIENT, canceloff);
772
773                 flags = ns_connect_lru_resize(ns) ?
774                         LDLM_CANCEL_LRUR : LDLM_CANCEL_AGED;
775                 to_free = !ns_connect_lru_resize(ns) &&
776                           opc == LDLM_ENQUEUE ? 1 : 0;
777
778                 /* Cancel LRU locks here _only_ if the server supports
779                  * EARLY_CANCEL. Otherwise we have to send extra CANCEL
780                  * RPC, which will make us slower. */
781                 if (avail > count)
782                         count += ldlm_cancel_lru_local(ns, cancels, to_free,
783                                                        avail - count, 0, flags);
784                 if (avail > count)
785                         pack = count;
786                 else
787                         pack = avail;
788                 req_capsule_set_size(pill, &RMF_DLM_REQ, RCL_CLIENT,
789                                      ldlm_request_bufsize(pack, opc));
790         }
791
792         rc = ptlrpc_request_pack(req, version, opc);
793         if (rc) {
794                 ldlm_lock_list_put(cancels, l_bl_ast, count);
795                 RETURN(rc);
796         }
797
798         if (ns_connect_cancelset(ns)) {
799                 if (canceloff) {
800                         dlm = req_capsule_client_get(pill, &RMF_DLM_REQ);
801                         LASSERT(dlm);
802                         /* Skip first lock handler in ldlm_request_pack(),
803                          * this method will incrment @lock_count according
804                          * to the lock handle amount actually written to
805                          * the buffer. */
806                         dlm->lock_count = canceloff;
807                 }
808                 /* Pack into the request @pack lock handles. */
809                 ldlm_cli_cancel_list(cancels, pack, req, 0);
810                 /* Prepare and send separate cancel RPC for others. */
811                 ldlm_cli_cancel_list(cancels, count - pack, NULL, 0);
812         } else {
813                 ldlm_lock_list_put(cancels, l_bl_ast, count);
814         }
815         RETURN(0);
816 }
817 EXPORT_SYMBOL(ldlm_prep_elc_req);
818
819 int ldlm_prep_enqueue_req(struct obd_export *exp, struct ptlrpc_request *req,
820                           cfs_list_t *cancels, int count)
821 {
822         return ldlm_prep_elc_req(exp, req, LUSTRE_DLM_VERSION, LDLM_ENQUEUE,
823                                  LDLM_ENQUEUE_CANCEL_OFF, cancels, count);
824 }
825 EXPORT_SYMBOL(ldlm_prep_enqueue_req);
826
827 struct ptlrpc_request *ldlm_enqueue_pack(struct obd_export *exp, int lvb_len)
828 {
829         struct ptlrpc_request *req;
830         int rc;
831         ENTRY;
832
833         req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_LDLM_ENQUEUE);
834         if (req == NULL)
835                 RETURN(ERR_PTR(-ENOMEM));
836
837         rc = ldlm_prep_enqueue_req(exp, req, NULL, 0);
838         if (rc) {
839                 ptlrpc_request_free(req);
840                 RETURN(ERR_PTR(rc));
841         }
842
843         req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER, lvb_len);
844         ptlrpc_request_set_replen(req);
845         RETURN(req);
846 }
847 EXPORT_SYMBOL(ldlm_enqueue_pack);
848
849 /**
850  * Client-side lock enqueue.
851  *
852  * If a request has some specific initialisation it is passed in \a reqp,
853  * otherwise it is created in ldlm_cli_enqueue.
854  *
855  * Supports sync and async requests, pass \a async flag accordingly. If a
856  * request was created in ldlm_cli_enqueue and it is the async request,
857  * pass it to the caller in \a reqp.
858  */
859 int ldlm_cli_enqueue(struct obd_export *exp, struct ptlrpc_request **reqp,
860                      struct ldlm_enqueue_info *einfo,
861                      const struct ldlm_res_id *res_id,
862                      ldlm_policy_data_t const *policy, __u64 *flags,
863                      void *lvb, __u32 lvb_len, enum lvb_type lvb_type,
864                      struct lustre_handle *lockh, int async)
865 {
866         struct ldlm_namespace *ns;
867         struct ldlm_lock      *lock;
868         struct ldlm_request   *body;
869         int                    is_replay = *flags & LDLM_FL_REPLAY;
870         int                    req_passed_in = 1;
871         int                    rc, err;
872         struct ptlrpc_request *req;
873         ENTRY;
874
875         LASSERT(exp != NULL);
876
877         ns = exp->exp_obd->obd_namespace;
878
879         /* If we're replaying this lock, just check some invariants.
880          * If we're creating a new lock, get everything all setup nice. */
881         if (is_replay) {
882                 lock = ldlm_handle2lock_long(lockh, 0);
883                 LASSERT(lock != NULL);
884                 LDLM_DEBUG(lock, "client-side enqueue START");
885                 LASSERT(exp == lock->l_conn_export);
886         } else {
887                 const struct ldlm_callback_suite cbs = {
888                         .lcs_completion = einfo->ei_cb_cp,
889                         .lcs_blocking   = einfo->ei_cb_bl,
890                         .lcs_glimpse    = einfo->ei_cb_gl
891                 };
892                 lock = ldlm_lock_create(ns, res_id, einfo->ei_type,
893                                         einfo->ei_mode, &cbs, einfo->ei_cbdata,
894                                         lvb_len, lvb_type);
895                 if (lock == NULL)
896                         RETURN(-ENOMEM);
897                 /* for the local lock, add the reference */
898                 ldlm_lock_addref_internal(lock, einfo->ei_mode);
899                 ldlm_lock2handle(lock, lockh);
900                 if (policy != NULL) {
901                         /* INODEBITS_INTEROP: If the server does not support
902                          * inodebits, we will request a plain lock in the
903                          * descriptor (ldlm_lock2desc() below) but use an
904                          * inodebits lock internally with both bits set.
905                          */
906                         if (einfo->ei_type == LDLM_IBITS &&
907                             !(exp_connect_flags(exp) &
908                               OBD_CONNECT_IBITS))
909                                 lock->l_policy_data.l_inodebits.bits =
910                                         MDS_INODELOCK_LOOKUP |
911                                         MDS_INODELOCK_UPDATE;
912                         else
913                                 lock->l_policy_data = *policy;
914                 }
915
916                 if (einfo->ei_type == LDLM_EXTENT)
917                         lock->l_req_extent = policy->l_extent;
918                 LDLM_DEBUG(lock, "client-side enqueue START, flags %llx\n",
919                            *flags);
920         }
921
922         lock->l_conn_export = exp;
923         lock->l_export = NULL;
924         lock->l_blocking_ast = einfo->ei_cb_bl;
925         lock->l_flags |= (*flags & (LDLM_FL_NO_LRU | LDLM_FL_EXCL));
926
927         /* lock not sent to server yet */
928
929         if (reqp == NULL || *reqp == NULL) {
930                 req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
931                                                 &RQF_LDLM_ENQUEUE,
932                                                 LUSTRE_DLM_VERSION,
933                                                 LDLM_ENQUEUE);
934                 if (req == NULL) {
935                         failed_lock_cleanup(ns, lock, einfo->ei_mode);
936                         LDLM_LOCK_RELEASE(lock);
937                         RETURN(-ENOMEM);
938                 }
939                 req_passed_in = 0;
940                 if (reqp)
941                         *reqp = req;
942         } else {
943                 int len;
944
945                 req = *reqp;
946                 len = req_capsule_get_size(&req->rq_pill, &RMF_DLM_REQ,
947                                            RCL_CLIENT);
948                 LASSERTF(len >= sizeof(*body), "buflen[%d] = %d, not %d\n",
949                          DLM_LOCKREQ_OFF, len, (int)sizeof(*body));
950         }
951
952         /* Dump lock data into the request buffer */
953         body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
954         ldlm_lock2desc(lock, &body->lock_desc);
955         body->lock_flags = ldlm_flags_to_wire(*flags);
956         body->lock_handle[0] = *lockh;
957
958         /* Continue as normal. */
959         if (!req_passed_in) {
960                 if (lvb_len > 0)
961                         req_capsule_extend(&req->rq_pill,
962                                            &RQF_LDLM_ENQUEUE_LVB);
963                 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
964                                      lvb_len);
965                 ptlrpc_request_set_replen(req);
966         }
967
968         /*
969          * Liblustre client doesn't get extent locks, except for O_APPEND case
970          * where [0, OBD_OBJECT_EOF] lock is taken, or truncate, where
971          * [i_size, OBD_OBJECT_EOF] lock is taken.
972          */
973         LASSERT(ergo(LIBLUSTRE_CLIENT, einfo->ei_type != LDLM_EXTENT ||
974                      policy->l_extent.end == OBD_OBJECT_EOF));
975
976         if (async) {
977                 LASSERT(reqp != NULL);
978                 RETURN(0);
979         }
980
981         LDLM_DEBUG(lock, "sending request");
982
983         rc = ptlrpc_queue_wait(req);
984
985         err = ldlm_cli_enqueue_fini(exp, req, einfo->ei_type, policy ? 1 : 0,
986                                     einfo->ei_mode, flags, lvb, lvb_len,
987                                     lockh, rc);
988
989         /* If ldlm_cli_enqueue_fini did not find the lock, we need to free
990          * one reference that we took */
991         if (err == -ENOLCK)
992                 LDLM_LOCK_RELEASE(lock);
993         else
994                 rc = err;
995
996         if (!req_passed_in && req != NULL) {
997                 ptlrpc_req_finished(req);
998                 if (reqp)
999                         *reqp = NULL;
1000         }
1001
1002         RETURN(rc);
1003 }
1004 EXPORT_SYMBOL(ldlm_cli_enqueue);
1005
1006 static int ldlm_cli_convert_local(struct ldlm_lock *lock, int new_mode,
1007                                   __u32 *flags)
1008 {
1009         struct ldlm_resource *res;
1010         int rc;
1011         ENTRY;
1012         if (ns_is_client(ldlm_lock_to_ns(lock))) {
1013                 CERROR("Trying to cancel local lock\n");
1014                 LBUG();
1015         }
1016         LDLM_DEBUG(lock, "client-side local convert");
1017
1018         res = ldlm_lock_convert(lock, new_mode, flags);
1019         if (res) {
1020                 ldlm_reprocess_all(res);
1021                 rc = 0;
1022         } else {
1023                 rc = LUSTRE_EDEADLK;
1024         }
1025         LDLM_DEBUG(lock, "client-side local convert handler END");
1026         LDLM_LOCK_PUT(lock);
1027         RETURN(rc);
1028 }
1029
1030 /* FIXME: one of ldlm_cli_convert or the server side should reject attempted
1031  * conversion of locks which are on the waiting or converting queue */
1032 /* Caller of this code is supposed to take care of lock readers/writers
1033    accounting */
1034 int ldlm_cli_convert(struct lustre_handle *lockh, int new_mode, __u32 *flags)
1035 {
1036         struct ldlm_request   *body;
1037         struct ldlm_reply     *reply;
1038         struct ldlm_lock      *lock;
1039         struct ldlm_resource  *res;
1040         struct ptlrpc_request *req;
1041         int                    rc;
1042         ENTRY;
1043
1044         lock = ldlm_handle2lock(lockh);
1045         if (!lock) {
1046                 LBUG();
1047                 RETURN(-EINVAL);
1048         }
1049         *flags = 0;
1050
1051         if (lock->l_conn_export == NULL)
1052                 RETURN(ldlm_cli_convert_local(lock, new_mode, flags));
1053
1054         LDLM_DEBUG(lock, "client-side convert");
1055
1056         req = ptlrpc_request_alloc_pack(class_exp2cliimp(lock->l_conn_export),
1057                                         &RQF_LDLM_CONVERT, LUSTRE_DLM_VERSION,
1058                                         LDLM_CONVERT);
1059         if (req == NULL) {
1060                 LDLM_LOCK_PUT(lock);
1061                 RETURN(-ENOMEM);
1062         }
1063
1064         body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1065         body->lock_handle[0] = lock->l_remote_handle;
1066
1067         body->lock_desc.l_req_mode = new_mode;
1068         body->lock_flags = ldlm_flags_to_wire(*flags);
1069
1070
1071         ptlrpc_request_set_replen(req);
1072         rc = ptlrpc_queue_wait(req);
1073         if (rc != ELDLM_OK)
1074                 GOTO(out, rc);
1075
1076         reply = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
1077         if (reply == NULL)
1078                 GOTO(out, rc = -EPROTO);
1079
1080         if (req->rq_status)
1081                 GOTO(out, rc = req->rq_status);
1082
1083         res = ldlm_lock_convert(lock, new_mode, &reply->lock_flags);
1084         if (res != NULL) {
1085                 ldlm_reprocess_all(res);
1086                 /* Go to sleep until the lock is granted. */
1087                 /* FIXME: or cancelled. */
1088                 if (lock->l_completion_ast) {
1089                         rc = lock->l_completion_ast(lock, LDLM_FL_WAIT_NOREPROC,
1090                                                     NULL);
1091                         if (rc)
1092                                 GOTO(out, rc);
1093                 }
1094         } else {
1095                 rc = LUSTRE_EDEADLK;
1096         }
1097         EXIT;
1098  out:
1099         LDLM_LOCK_PUT(lock);
1100         ptlrpc_req_finished(req);
1101         return rc;
1102 }
1103 EXPORT_SYMBOL(ldlm_cli_convert);
1104
1105 /**
1106  * Cancel locks locally.
1107  * Returns:
1108  * \retval LDLM_FL_LOCAL_ONLY if there is no need for a CANCEL RPC to the server
1109  * \retval LDLM_FL_CANCELING otherwise;
1110  * \retval LDLM_FL_BL_AST if there is a need for a separate CANCEL RPC.
1111  */
1112 static __u64 ldlm_cli_cancel_local(struct ldlm_lock *lock)
1113 {
1114         __u64 rc = LDLM_FL_LOCAL_ONLY;
1115         ENTRY;
1116
1117         if (lock->l_conn_export) {
1118                 bool local_only;
1119
1120                 LDLM_DEBUG(lock, "client-side cancel");
1121                 /* Set this flag to prevent others from getting new references*/
1122                 lock_res_and_lock(lock);
1123                 lock->l_flags |= LDLM_FL_CBPENDING;
1124                 local_only = !!(lock->l_flags &
1125                                 (LDLM_FL_LOCAL_ONLY|LDLM_FL_CANCEL_ON_BLOCK));
1126                 ldlm_cancel_callback(lock);
1127                 rc = (lock->l_flags & LDLM_FL_BL_AST) ?
1128                         LDLM_FL_BL_AST : LDLM_FL_CANCELING;
1129                 unlock_res_and_lock(lock);
1130
1131                 if (local_only) {
1132                         CDEBUG(D_DLMTRACE, "not sending request (at caller's "
1133                                "instruction)\n");
1134                         rc = LDLM_FL_LOCAL_ONLY;
1135                 }
1136                 ldlm_lock_cancel(lock);
1137         } else {
1138                 if (ns_is_client(ldlm_lock_to_ns(lock))) {
1139                         LDLM_ERROR(lock, "Trying to cancel local lock");
1140                         LBUG();
1141                 }
1142                 LDLM_DEBUG(lock, "server-side local cancel");
1143                 ldlm_lock_cancel(lock);
1144                 ldlm_reprocess_all(lock->l_resource);
1145         }
1146
1147         RETURN(rc);
1148 }
1149
1150 /**
1151  * Pack \a count locks in \a head into ldlm_request buffer of request \a req.
1152  */
1153 static void ldlm_cancel_pack(struct ptlrpc_request *req,
1154                              cfs_list_t *head, int count)
1155 {
1156         struct ldlm_request *dlm;
1157         struct ldlm_lock *lock;
1158         int max, packed = 0;
1159         ENTRY;
1160
1161         dlm = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1162         LASSERT(dlm != NULL);
1163
1164         /* Check the room in the request buffer. */
1165         max = req_capsule_get_size(&req->rq_pill, &RMF_DLM_REQ, RCL_CLIENT) -
1166                 sizeof(struct ldlm_request);
1167         max /= sizeof(struct lustre_handle);
1168         max += LDLM_LOCKREQ_HANDLES;
1169         LASSERT(max >= dlm->lock_count + count);
1170
1171         /* XXX: it would be better to pack lock handles grouped by resource.
1172          * so that the server cancel would call filter_lvbo_update() less
1173          * frequently. */
1174         cfs_list_for_each_entry(lock, head, l_bl_ast) {
1175                 if (!count--)
1176                         break;
1177                 LASSERT(lock->l_conn_export);
1178                 /* Pack the lock handle to the given request buffer. */
1179                 LDLM_DEBUG(lock, "packing");
1180                 dlm->lock_handle[dlm->lock_count++] = lock->l_remote_handle;
1181                 packed++;
1182         }
1183         CDEBUG(D_DLMTRACE, "%d locks packed\n", packed);
1184         EXIT;
1185 }
1186
1187 /**
1188  * Prepare and send a batched cancel RPC. It will include \a count lock
1189  * handles of locks given in \a cancels list. */
1190 int ldlm_cli_cancel_req(struct obd_export *exp, cfs_list_t *cancels,
1191                         int count, ldlm_cancel_flags_t flags)
1192 {
1193         struct ptlrpc_request *req = NULL;
1194         struct obd_import *imp;
1195         int free, sent = 0;
1196         int rc = 0;
1197         ENTRY;
1198
1199         LASSERT(exp != NULL);
1200         LASSERT(count > 0);
1201
1202         CFS_FAIL_TIMEOUT(OBD_FAIL_LDLM_PAUSE_CANCEL, cfs_fail_val);
1203
1204         if (CFS_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_RACE))
1205                 RETURN(count);
1206
1207         free = ldlm_format_handles_avail(class_exp2cliimp(exp),
1208                                          &RQF_LDLM_CANCEL, RCL_CLIENT, 0);
1209         if (count > free)
1210                 count = free;
1211
1212         while (1) {
1213                 imp = class_exp2cliimp(exp);
1214                 if (imp == NULL || imp->imp_invalid) {
1215                         CDEBUG(D_DLMTRACE,
1216                                "skipping cancel on invalid import %p\n", imp);
1217                         RETURN(count);
1218                 }
1219
1220                 req = ptlrpc_request_alloc(imp, &RQF_LDLM_CANCEL);
1221                 if (req == NULL)
1222                         GOTO(out, rc = -ENOMEM);
1223
1224                 req_capsule_filled_sizes(&req->rq_pill, RCL_CLIENT);
1225                 req_capsule_set_size(&req->rq_pill, &RMF_DLM_REQ, RCL_CLIENT,
1226                                      ldlm_request_bufsize(count, LDLM_CANCEL));
1227
1228                 rc = ptlrpc_request_pack(req, LUSTRE_DLM_VERSION, LDLM_CANCEL);
1229                 if (rc) {
1230                         ptlrpc_request_free(req);
1231                         GOTO(out, rc);
1232                 }
1233
1234                 req->rq_request_portal = LDLM_CANCEL_REQUEST_PORTAL;
1235                 req->rq_reply_portal = LDLM_CANCEL_REPLY_PORTAL;
1236                 ptlrpc_at_set_req_timeout(req);
1237
1238                 ldlm_cancel_pack(req, cancels, count);
1239
1240                 ptlrpc_request_set_replen(req);
1241                 if (flags & LCF_ASYNC) {
1242                         ptlrpcd_add_req(req, PDL_POLICY_LOCAL, -1);
1243                         sent = count;
1244                         GOTO(out, 0);
1245                 } else {
1246                         rc = ptlrpc_queue_wait(req);
1247                 }
1248                 if (rc == LUSTRE_ESTALE) {
1249                         CDEBUG(D_DLMTRACE, "client/server (nid %s) "
1250                                "out of sync -- not fatal\n",
1251                                libcfs_nid2str(req->rq_import->
1252                                               imp_connection->c_peer.nid));
1253                         rc = 0;
1254                 } else if (rc == -ETIMEDOUT && /* check there was no reconnect*/
1255                            req->rq_import_generation == imp->imp_generation) {
1256                         ptlrpc_req_finished(req);
1257                         continue;
1258                 } else if (rc != ELDLM_OK) {
1259                         /* -ESHUTDOWN is common on umount */
1260                         CDEBUG_LIMIT(rc == -ESHUTDOWN ? D_DLMTRACE : D_ERROR,
1261                                      "Got rc %d from cancel RPC: "
1262                                      "canceling anyway\n", rc);
1263                         break;
1264                 }
1265                 sent = count;
1266                 break;
1267         }
1268
1269         ptlrpc_req_finished(req);
1270         EXIT;
1271 out:
1272         return sent ? sent : rc;
1273 }
1274 EXPORT_SYMBOL(ldlm_cli_cancel_req);
1275
1276 static inline struct ldlm_pool *ldlm_imp2pl(struct obd_import *imp)
1277 {
1278         LASSERT(imp != NULL);
1279         return &imp->imp_obd->obd_namespace->ns_pool;
1280 }
1281
1282 /**
1283  * Update client's OBD pool related fields with new SLV and Limit from \a req.
1284  */
1285 int ldlm_cli_update_pool(struct ptlrpc_request *req)
1286 {
1287         struct obd_device *obd;
1288         __u64 new_slv;
1289         __u32 new_limit;
1290         ENTRY;
1291         if (unlikely(!req->rq_import || !req->rq_import->imp_obd ||
1292                      !imp_connect_lru_resize(req->rq_import)))
1293         {
1294                 /*
1295                  * Do nothing for corner cases.
1296                  */
1297                 RETURN(0);
1298         }
1299
1300         /* In some cases RPC may contain SLV and limit zeroed out. This
1301          * is the case when server does not support LRU resize feature.
1302          * This is also possible in some recovery cases when server-side
1303          * reqs have no reference to the OBD export and thus access to
1304          * server-side namespace is not possible. */
1305         if (lustre_msg_get_slv(req->rq_repmsg) == 0 ||
1306             lustre_msg_get_limit(req->rq_repmsg) == 0) {
1307                 DEBUG_REQ(D_HA, req, "Zero SLV or Limit found "
1308                           "(SLV: "LPU64", Limit: %u)",
1309                           lustre_msg_get_slv(req->rq_repmsg),
1310                           lustre_msg_get_limit(req->rq_repmsg));
1311                 RETURN(0);
1312         }
1313
1314         new_limit = lustre_msg_get_limit(req->rq_repmsg);
1315         new_slv = lustre_msg_get_slv(req->rq_repmsg);
1316         obd = req->rq_import->imp_obd;
1317
1318         /* Set new SLV and limit in OBD fields to make them accessible
1319          * to the pool thread. We do not access obd_namespace and pool
1320          * directly here as there is no reliable way to make sure that
1321          * they are still alive at cleanup time. Evil races are possible
1322          * which may cause Oops at that time. */
1323         write_lock(&obd->obd_pool_lock);
1324         obd->obd_pool_slv = new_slv;
1325         obd->obd_pool_limit = new_limit;
1326         write_unlock(&obd->obd_pool_lock);
1327
1328         RETURN(0);
1329 }
1330 EXPORT_SYMBOL(ldlm_cli_update_pool);
1331
1332 /**
1333  * Client side lock cancel.
1334  *
1335  * Lock must not have any readers or writers by this time.
1336  */
1337 int ldlm_cli_cancel(struct lustre_handle *lockh,
1338                     ldlm_cancel_flags_t cancel_flags)
1339 {
1340         struct obd_export *exp;
1341         int avail, flags, count = 1;
1342         __u64 rc = 0;
1343         struct ldlm_namespace *ns;
1344         struct ldlm_lock *lock;
1345         CFS_LIST_HEAD(cancels);
1346         ENTRY;
1347
1348         /* concurrent cancels on the same handle can happen */
1349         lock = ldlm_handle2lock_long(lockh, LDLM_FL_CANCELING);
1350         if (lock == NULL) {
1351                 LDLM_DEBUG_NOLOCK("lock is already being destroyed\n");
1352                 RETURN(0);
1353         }
1354
1355         rc = ldlm_cli_cancel_local(lock);
1356         if (rc == LDLM_FL_LOCAL_ONLY || cancel_flags & LCF_LOCAL) {
1357                 LDLM_LOCK_RELEASE(lock);
1358                 RETURN(0);
1359         }
1360         /* Even if the lock is marked as LDLM_FL_BL_AST, this is a LDLM_CANCEL
1361          * RPC which goes to canceld portal, so we can cancel other LRU locks
1362          * here and send them all as one LDLM_CANCEL RPC. */
1363         LASSERT(cfs_list_empty(&lock->l_bl_ast));
1364         cfs_list_add(&lock->l_bl_ast, &cancels);
1365
1366         exp = lock->l_conn_export;
1367         if (exp_connect_cancelset(exp)) {
1368                 avail = ldlm_format_handles_avail(class_exp2cliimp(exp),
1369                                                   &RQF_LDLM_CANCEL,
1370                                                   RCL_CLIENT, 0);
1371                 LASSERT(avail > 0);
1372
1373                 ns = ldlm_lock_to_ns(lock);
1374                 flags = ns_connect_lru_resize(ns) ?
1375                         LDLM_CANCEL_LRUR : LDLM_CANCEL_AGED;
1376                 count += ldlm_cancel_lru_local(ns, &cancels, 0, avail - 1,
1377                                                LCF_BL_AST, flags);
1378         }
1379         ldlm_cli_cancel_list(&cancels, count, NULL, cancel_flags);
1380         RETURN(0);
1381 }
1382 EXPORT_SYMBOL(ldlm_cli_cancel);
1383
1384 /**
1385  * Locally cancel up to \a count locks in list \a cancels.
1386  * Return the number of cancelled locks.
1387  */
1388 int ldlm_cli_cancel_list_local(cfs_list_t *cancels, int count,
1389                                ldlm_cancel_flags_t flags)
1390 {
1391         CFS_LIST_HEAD(head);
1392         struct ldlm_lock *lock, *next;
1393         int left = 0, bl_ast = 0;
1394         __u64 rc;
1395
1396         left = count;
1397         cfs_list_for_each_entry_safe(lock, next, cancels, l_bl_ast) {
1398                 if (left-- == 0)
1399                         break;
1400
1401                 if (flags & LCF_LOCAL) {
1402                         rc = LDLM_FL_LOCAL_ONLY;
1403                         ldlm_lock_cancel(lock);
1404                 } else {
1405                         rc = ldlm_cli_cancel_local(lock);
1406                 }
1407                 /* Until we have compound requests and can send LDLM_CANCEL
1408                  * requests batched with generic RPCs, we need to send cancels
1409                  * with the LDLM_FL_BL_AST flag in a separate RPC from
1410                  * the one being generated now. */
1411                 if (!(flags & LCF_BL_AST) && (rc == LDLM_FL_BL_AST)) {
1412                         LDLM_DEBUG(lock, "Cancel lock separately");
1413                         cfs_list_del_init(&lock->l_bl_ast);
1414                         cfs_list_add(&lock->l_bl_ast, &head);
1415                         bl_ast++;
1416                         continue;
1417                 }
1418                 if (rc == LDLM_FL_LOCAL_ONLY) {
1419                         /* CANCEL RPC should not be sent to server. */
1420                         cfs_list_del_init(&lock->l_bl_ast);
1421                         LDLM_LOCK_RELEASE(lock);
1422                         count--;
1423                 }
1424         }
1425         if (bl_ast > 0) {
1426                 count -= bl_ast;
1427                 ldlm_cli_cancel_list(&head, bl_ast, NULL, 0);
1428         }
1429
1430         RETURN(count);
1431 }
1432 EXPORT_SYMBOL(ldlm_cli_cancel_list_local);
1433
1434 /**
1435  * Cancel as many locks as possible w/o sending any RPCs (e.g. to write back
1436  * dirty data, to close a file, ...) or waiting for any RPCs in-flight (e.g.
1437  * readahead requests, ...)
1438  */
1439 static ldlm_policy_res_t ldlm_cancel_no_wait_policy(struct ldlm_namespace *ns,
1440                                                     struct ldlm_lock *lock,
1441                                                     int unused, int added,
1442                                                     int count)
1443 {
1444         ldlm_policy_res_t result = LDLM_POLICY_CANCEL_LOCK;
1445         ldlm_cancel_for_recovery cb = ns->ns_cancel_for_recovery;
1446         lock_res_and_lock(lock);
1447
1448         /* don't check added & count since we want to process all locks
1449          * from unused list */
1450         switch (lock->l_resource->lr_type) {
1451                 case LDLM_EXTENT:
1452                 case LDLM_IBITS:
1453                         if (cb && cb(lock))
1454                                 break;
1455                 default:
1456                         result = LDLM_POLICY_SKIP_LOCK;
1457                         lock->l_flags |= LDLM_FL_SKIPPED;
1458                         break;
1459         }
1460
1461         unlock_res_and_lock(lock);
1462         RETURN(result);
1463 }
1464
1465 /**
1466  * Callback function for LRU-resize policy. Decides whether to keep
1467  * \a lock in LRU for current \a LRU size \a unused, added in current
1468  * scan \a added and number of locks to be preferably canceled \a count.
1469  *
1470  * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1471  *
1472  * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1473  */
1474 static ldlm_policy_res_t ldlm_cancel_lrur_policy(struct ldlm_namespace *ns,
1475                                                  struct ldlm_lock *lock,
1476                                                  int unused, int added,
1477                                                  int count)
1478 {
1479         cfs_time_t cur = cfs_time_current();
1480         struct ldlm_pool *pl = &ns->ns_pool;
1481         __u64 slv, lvf, lv;
1482         cfs_time_t la;
1483
1484         /* Stop LRU processing when we reach past @count or have checked all
1485          * locks in LRU. */
1486         if (count && added >= count)
1487                 return LDLM_POLICY_KEEP_LOCK;
1488
1489         slv = ldlm_pool_get_slv(pl);
1490         lvf = ldlm_pool_get_lvf(pl);
1491         la = cfs_duration_sec(cfs_time_sub(cur,
1492                               lock->l_last_used));
1493         lv = lvf * la * unused;
1494
1495         /* Inform pool about current CLV to see it via proc. */
1496         ldlm_pool_set_clv(pl, lv);
1497
1498         /* Stop when SLV is not yet come from server or lv is smaller than
1499          * it is. */
1500         return (slv == 0 || lv < slv) ?
1501                 LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK;
1502 }
1503
1504 /**
1505  * Callback function for proc used policy. Makes decision whether to keep
1506  * \a lock in LRU for current \a LRU size \a unused, added in current scan \a
1507  * added and number of locks to be preferably canceled \a count.
1508  *
1509  * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1510  *
1511  * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1512  */
1513 static ldlm_policy_res_t ldlm_cancel_passed_policy(struct ldlm_namespace *ns,
1514                                                    struct ldlm_lock *lock,
1515                                                    int unused, int added,
1516                                                    int count)
1517 {
1518         /* Stop LRU processing when we reach past @count or have checked all
1519          * locks in LRU. */
1520         return (added >= count) ?
1521                 LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK;
1522 }
1523
1524 /**
1525  * Callback function for aged policy. Makes decision whether to keep \a lock in
1526  * LRU for current LRU size \a unused, added in current scan \a added and
1527  * number of locks to be preferably canceled \a count.
1528  *
1529  * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1530  *
1531  * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1532  */
1533 static ldlm_policy_res_t ldlm_cancel_aged_policy(struct ldlm_namespace *ns,
1534                                                  struct ldlm_lock *lock,
1535                                                  int unused, int added,
1536                                                  int count)
1537 {
1538         /* Stop LRU processing if young lock is found and we reach past count */
1539         return ((added >= count) &&
1540                 cfs_time_before(cfs_time_current(),
1541                                 cfs_time_add(lock->l_last_used,
1542                                              ns->ns_max_age))) ?
1543                 LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK;
1544 }
1545
1546 /**
1547  * Callback function for default policy. Makes decision whether to keep \a lock
1548  * in LRU for current LRU size \a unused, added in current scan \a added and
1549  * number of locks to be preferably canceled \a count.
1550  *
1551  * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1552  *
1553  * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1554  */
1555 static ldlm_policy_res_t ldlm_cancel_default_policy(struct ldlm_namespace *ns,
1556                                                     struct ldlm_lock *lock,
1557                                                     int unused, int added,
1558                                                     int count)
1559 {
1560         /* Stop LRU processing when we reach past count or have checked all
1561          * locks in LRU. */
1562         return (added >= count) ?
1563                 LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK;
1564 }
1565
1566 typedef ldlm_policy_res_t (*ldlm_cancel_lru_policy_t)(struct ldlm_namespace *,
1567                                                       struct ldlm_lock *, int,
1568                                                       int, int);
1569
1570 static ldlm_cancel_lru_policy_t
1571 ldlm_cancel_lru_policy(struct ldlm_namespace *ns, int flags)
1572 {
1573         if (flags & LDLM_CANCEL_NO_WAIT)
1574                 return ldlm_cancel_no_wait_policy;
1575
1576         if (ns_connect_lru_resize(ns)) {
1577                 if (flags & LDLM_CANCEL_SHRINK)
1578                         /* We kill passed number of old locks. */
1579                         return ldlm_cancel_passed_policy;
1580                 else if (flags & LDLM_CANCEL_LRUR)
1581                         return ldlm_cancel_lrur_policy;
1582                 else if (flags & LDLM_CANCEL_PASSED)
1583                         return ldlm_cancel_passed_policy;
1584         } else {
1585                 if (flags & LDLM_CANCEL_AGED)
1586                         return ldlm_cancel_aged_policy;
1587         }
1588
1589         return ldlm_cancel_default_policy;
1590 }
1591
1592 /**
1593  * - Free space in LRU for \a count new locks,
1594  *   redundant unused locks are canceled locally;
1595  * - also cancel locally unused aged locks;
1596  * - do not cancel more than \a max locks;
1597  * - GET the found locks and add them into the \a cancels list.
1598  *
1599  * A client lock can be added to the l_bl_ast list only when it is
1600  * marked LDLM_FL_CANCELING. Otherwise, somebody is already doing
1601  * CANCEL.  There are the following use cases:
1602  * ldlm_cancel_resource_local(), ldlm_cancel_lru_local() and
1603  * ldlm_cli_cancel(), which check and set this flag properly. As any
1604  * attempt to cancel a lock rely on this flag, l_bl_ast list is accessed
1605  * later without any special locking.
1606  *
1607  * Calling policies for enabled LRU resize:
1608  * ----------------------------------------
1609  * flags & LDLM_CANCEL_LRUR - use LRU resize policy (SLV from server) to
1610  *                            cancel not more than \a count locks;
1611  *
1612  * flags & LDLM_CANCEL_PASSED - cancel \a count number of old locks (located at
1613  *                              the beginning of LRU list);
1614  *
1615  * flags & LDLM_CANCEL_SHRINK - cancel not more than \a count locks according to
1616  *                              memory pressre policy function;
1617  *
1618  * flags & LDLM_CANCEL_AGED - cancel \a count locks according to "aged policy".
1619  *
1620  * flags & LDLM_CANCEL_NO_WAIT - cancel as many unused locks as possible
1621  *                               (typically before replaying locks) w/o
1622  *                               sending any RPCs or waiting for any
1623  *                               outstanding RPC to complete.
1624  */
1625 static int ldlm_prepare_lru_list(struct ldlm_namespace *ns, cfs_list_t *cancels,
1626                                  int count, int max, int flags)
1627 {
1628         ldlm_cancel_lru_policy_t pf;
1629         struct ldlm_lock *lock, *next;
1630         int added = 0, unused, remained;
1631         ENTRY;
1632
1633         spin_lock(&ns->ns_lock);
1634         unused = ns->ns_nr_unused;
1635         remained = unused;
1636
1637         if (!ns_connect_lru_resize(ns))
1638                 count += unused - ns->ns_max_unused;
1639
1640         pf = ldlm_cancel_lru_policy(ns, flags);
1641         LASSERT(pf != NULL);
1642
1643         while (!cfs_list_empty(&ns->ns_unused_list)) {
1644                 ldlm_policy_res_t result;
1645
1646                 /* all unused locks */
1647                 if (remained-- <= 0)
1648                         break;
1649
1650                 /* For any flags, stop scanning if @max is reached. */
1651                 if (max && added >= max)
1652                         break;
1653
1654                 cfs_list_for_each_entry_safe(lock, next, &ns->ns_unused_list,
1655                                              l_lru) {
1656                         /* No locks which got blocking requests. */
1657                         LASSERT(!(lock->l_flags & LDLM_FL_BL_AST));
1658
1659                         if (flags & LDLM_CANCEL_NO_WAIT &&
1660                             lock->l_flags & LDLM_FL_SKIPPED)
1661                                 /* already processed */
1662                                 continue;
1663
1664                         /* Somebody is already doing CANCEL. No need for this
1665                          * lock in LRU, do not traverse it again. */
1666                         if (!(lock->l_flags & LDLM_FL_CANCELING))
1667                                 break;
1668
1669                         ldlm_lock_remove_from_lru_nolock(lock);
1670                 }
1671                 if (&lock->l_lru == &ns->ns_unused_list)
1672                         break;
1673
1674                 LDLM_LOCK_GET(lock);
1675                 spin_unlock(&ns->ns_lock);
1676                 lu_ref_add(&lock->l_reference, __FUNCTION__, cfs_current());
1677
1678                 /* Pass the lock through the policy filter and see if it
1679                  * should stay in LRU.
1680                  *
1681                  * Even for shrinker policy we stop scanning if
1682                  * we find a lock that should stay in the cache.
1683                  * We should take into account lock age anyway
1684                  * as a new lock is a valuable resource even if
1685                  * it has a low weight.
1686                  *
1687                  * That is, for shrinker policy we drop only
1688                  * old locks, but additionally choose them by
1689                  * their weight. Big extent locks will stay in
1690                  * the cache. */
1691                 result = pf(ns, lock, unused, added, count);
1692                 if (result == LDLM_POLICY_KEEP_LOCK) {
1693                         lu_ref_del(&lock->l_reference,
1694                                    __FUNCTION__, cfs_current());
1695                         LDLM_LOCK_RELEASE(lock);
1696                         spin_lock(&ns->ns_lock);
1697                         break;
1698                 }
1699                 if (result == LDLM_POLICY_SKIP_LOCK) {
1700                         lu_ref_del(&lock->l_reference,
1701                                    __func__, cfs_current());
1702                         LDLM_LOCK_RELEASE(lock);
1703                         spin_lock(&ns->ns_lock);
1704                         continue;
1705                 }
1706
1707                 lock_res_and_lock(lock);
1708                 /* Check flags again under the lock. */
1709                 if ((lock->l_flags & LDLM_FL_CANCELING) ||
1710                     (ldlm_lock_remove_from_lru(lock) == 0)) {
1711                         /* Another thread is removing lock from LRU, or
1712                          * somebody is already doing CANCEL, or there
1713                          * is a blocking request which will send cancel
1714                          * by itself, or the lock is no longer unused. */
1715                         unlock_res_and_lock(lock);
1716                         lu_ref_del(&lock->l_reference,
1717                                    __FUNCTION__, cfs_current());
1718                         LDLM_LOCK_RELEASE(lock);
1719                         spin_lock(&ns->ns_lock);
1720                         continue;
1721                 }
1722                 LASSERT(!lock->l_readers && !lock->l_writers);
1723
1724                 /* If we have chosen to cancel this lock voluntarily, we
1725                  * better send cancel notification to server, so that it
1726                  * frees appropriate state. This might lead to a race
1727                  * where while we are doing cancel here, server is also
1728                  * silently cancelling this lock. */
1729                 lock->l_flags &= ~LDLM_FL_CANCEL_ON_BLOCK;
1730
1731                 /* Setting the CBPENDING flag is a little misleading,
1732                  * but prevents an important race; namely, once
1733                  * CBPENDING is set, the lock can accumulate no more
1734                  * readers/writers. Since readers and writers are
1735                  * already zero here, ldlm_lock_decref() won't see
1736                  * this flag and call l_blocking_ast */
1737                 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_CANCELING;
1738
1739                 /* We can't re-add to l_lru as it confuses the
1740                  * refcounting in ldlm_lock_remove_from_lru() if an AST
1741                  * arrives after we drop lr_lock below. We use l_bl_ast
1742                  * and can't use l_pending_chain as it is used both on
1743                  * server and client nevertheless bug 5666 says it is
1744                  * used only on server */
1745                 LASSERT(cfs_list_empty(&lock->l_bl_ast));
1746                 cfs_list_add(&lock->l_bl_ast, cancels);
1747                 unlock_res_and_lock(lock);
1748                 lu_ref_del(&lock->l_reference, __FUNCTION__, cfs_current());
1749                 spin_lock(&ns->ns_lock);
1750                 added++;
1751                 unused--;
1752         }
1753         spin_unlock(&ns->ns_lock);
1754         RETURN(added);
1755 }
1756
1757 int ldlm_cancel_lru_local(struct ldlm_namespace *ns, cfs_list_t *cancels,
1758                           int count, int max, ldlm_cancel_flags_t cancel_flags,
1759                           int flags)
1760 {
1761         int added;
1762         added = ldlm_prepare_lru_list(ns, cancels, count, max, flags);
1763         if (added <= 0)
1764                 return added;
1765         return ldlm_cli_cancel_list_local(cancels, added, cancel_flags);
1766 }
1767
1768 /**
1769  * Cancel at least \a nr locks from given namespace LRU.
1770  *
1771  * When called with LCF_ASYNC the blocking callback will be handled
1772  * in a thread and this function will return after the thread has been
1773  * asked to call the callback.  When called with LCF_ASYNC the blocking
1774  * callback will be performed in this function.
1775  */
1776 int ldlm_cancel_lru(struct ldlm_namespace *ns, int nr,
1777                     ldlm_cancel_flags_t cancel_flags,
1778                     int flags)
1779 {
1780         CFS_LIST_HEAD(cancels);
1781         int count, rc;
1782         ENTRY;
1783
1784 #ifndef __KERNEL__
1785         cancel_flags &= ~LCF_ASYNC; /* force to be sync in user space */
1786 #endif
1787         /* Just prepare the list of locks, do not actually cancel them yet.
1788          * Locks are cancelled later in a separate thread. */
1789         count = ldlm_prepare_lru_list(ns, &cancels, nr, 0, flags);
1790         rc = ldlm_bl_to_thread_list(ns, NULL, &cancels, count, cancel_flags);
1791         if (rc == 0)
1792                 RETURN(count);
1793
1794         RETURN(0);
1795 }
1796
1797 /**
1798  * Find and cancel locally unused locks found on resource, matched to the
1799  * given policy, mode. GET the found locks and add them into the \a cancels
1800  * list.
1801  */
1802 int ldlm_cancel_resource_local(struct ldlm_resource *res,
1803                                cfs_list_t *cancels,
1804                                ldlm_policy_data_t *policy,
1805                                ldlm_mode_t mode, int lock_flags,
1806                                ldlm_cancel_flags_t cancel_flags, void *opaque)
1807 {
1808         struct ldlm_lock *lock;
1809         int count = 0;
1810         ENTRY;
1811
1812         lock_res(res);
1813         cfs_list_for_each_entry(lock, &res->lr_granted, l_res_link) {
1814                 if (opaque != NULL && lock->l_ast_data != opaque) {
1815                         LDLM_ERROR(lock, "data %p doesn't match opaque %p",
1816                                    lock->l_ast_data, opaque);
1817                         //LBUG();
1818                         continue;
1819                 }
1820
1821                 if (lock->l_readers || lock->l_writers)
1822                         continue;
1823
1824                 /* If somebody is already doing CANCEL, or blocking AST came,
1825                  * skip this lock. */
1826                 if (lock->l_flags & LDLM_FL_BL_AST ||
1827                     lock->l_flags & LDLM_FL_CANCELING)
1828                         continue;
1829
1830                 if (lockmode_compat(lock->l_granted_mode, mode))
1831                         continue;
1832
1833                 /* If policy is given and this is IBITS lock, add to list only
1834                  * those locks that match by policy. */
1835                 if (policy && (lock->l_resource->lr_type == LDLM_IBITS) &&
1836                     !(lock->l_policy_data.l_inodebits.bits &
1837                       policy->l_inodebits.bits))
1838                         continue;
1839
1840                 /* See CBPENDING comment in ldlm_cancel_lru */
1841                 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_CANCELING |
1842                                  lock_flags;
1843
1844                 LASSERT(cfs_list_empty(&lock->l_bl_ast));
1845                 cfs_list_add(&lock->l_bl_ast, cancels);
1846                 LDLM_LOCK_GET(lock);
1847                 count++;
1848         }
1849         unlock_res(res);
1850
1851         RETURN(ldlm_cli_cancel_list_local(cancels, count, cancel_flags));
1852 }
1853 EXPORT_SYMBOL(ldlm_cancel_resource_local);
1854
1855 /**
1856  * Cancel client-side locks from a list and send/prepare cancel RPCs to the
1857  * server.
1858  * If \a req is NULL, send CANCEL request to server with handles of locks
1859  * in the \a cancels. If EARLY_CANCEL is not supported, send CANCEL requests
1860  * separately per lock.
1861  * If \a req is not NULL, put handles of locks in \a cancels into the request
1862  * buffer at the offset \a off.
1863  * Destroy \a cancels at the end.
1864  */
1865 int ldlm_cli_cancel_list(cfs_list_t *cancels, int count,
1866                          struct ptlrpc_request *req, ldlm_cancel_flags_t flags)
1867 {
1868         struct ldlm_lock *lock;
1869         int res = 0;
1870         ENTRY;
1871
1872         if (cfs_list_empty(cancels) || count == 0)
1873                 RETURN(0);
1874
1875         /* XXX: requests (both batched and not) could be sent in parallel.
1876          * Usually it is enough to have just 1 RPC, but it is possible that
1877          * there are too many locks to be cancelled in LRU or on a resource.
1878          * It would also speed up the case when the server does not support
1879          * the feature. */
1880         while (count > 0) {
1881                 LASSERT(!cfs_list_empty(cancels));
1882                 lock = cfs_list_entry(cancels->next, struct ldlm_lock,
1883                                       l_bl_ast);
1884                 LASSERT(lock->l_conn_export);
1885
1886                 if (exp_connect_cancelset(lock->l_conn_export)) {
1887                         res = count;
1888                         if (req)
1889                                 ldlm_cancel_pack(req, cancels, count);
1890                         else
1891                                 res = ldlm_cli_cancel_req(lock->l_conn_export,
1892                                                           cancels, count,
1893                                                           flags);
1894                 } else {
1895                         res = ldlm_cli_cancel_req(lock->l_conn_export,
1896                                                   cancels, 1, flags);
1897                 }
1898
1899                 if (res < 0) {
1900                         CDEBUG_LIMIT(res == -ESHUTDOWN ? D_DLMTRACE : D_ERROR,
1901                                      "ldlm_cli_cancel_list: %d\n", res);
1902                         res = count;
1903                 }
1904
1905                 count -= res;
1906                 ldlm_lock_list_put(cancels, l_bl_ast, res);
1907         }
1908         LASSERT(count == 0);
1909         RETURN(0);
1910 }
1911 EXPORT_SYMBOL(ldlm_cli_cancel_list);
1912
1913 /**
1914  * Cancel all locks on a resource that have 0 readers/writers.
1915  *
1916  * If flags & LDLM_FL_LOCAL_ONLY, throw the locks away without trying
1917  * to notify the server. */
1918 int ldlm_cli_cancel_unused_resource(struct ldlm_namespace *ns,
1919                                     const struct ldlm_res_id *res_id,
1920                                     ldlm_policy_data_t *policy,
1921                                     ldlm_mode_t mode,
1922                                     ldlm_cancel_flags_t flags,
1923                                     void *opaque)
1924 {
1925         struct ldlm_resource *res;
1926         CFS_LIST_HEAD(cancels);
1927         int count;
1928         int rc;
1929         ENTRY;
1930
1931         res = ldlm_resource_get(ns, NULL, res_id, 0, 0);
1932         if (res == NULL) {
1933                 /* This is not a problem. */
1934                 CDEBUG(D_INFO, "No resource "LPU64"\n", res_id->name[0]);
1935                 RETURN(0);
1936         }
1937
1938         LDLM_RESOURCE_ADDREF(res);
1939         count = ldlm_cancel_resource_local(res, &cancels, policy, mode,
1940                                            0, flags | LCF_BL_AST, opaque);
1941         rc = ldlm_cli_cancel_list(&cancels, count, NULL, flags);
1942         if (rc != ELDLM_OK)
1943                 CERROR("canceling unused lock "DLDLMRES": rc = %d\n",
1944                        PLDLMRES(res), rc);
1945
1946         LDLM_RESOURCE_DELREF(res);
1947         ldlm_resource_putref(res);
1948         RETURN(0);
1949 }
1950 EXPORT_SYMBOL(ldlm_cli_cancel_unused_resource);
1951
1952 struct ldlm_cli_cancel_arg {
1953         int     lc_flags;
1954         void   *lc_opaque;
1955 };
1956
1957 static int ldlm_cli_hash_cancel_unused(cfs_hash_t *hs, cfs_hash_bd_t *bd,
1958                                        cfs_hlist_node_t *hnode, void *arg)
1959 {
1960         struct ldlm_resource           *res = cfs_hash_object(hs, hnode);
1961         struct ldlm_cli_cancel_arg     *lc = arg;
1962
1963         ldlm_cli_cancel_unused_resource(ldlm_res_to_ns(res), &res->lr_name,
1964                                         NULL, LCK_MINMODE, lc->lc_flags,
1965                                         lc->lc_opaque);
1966         /* must return 0 for hash iteration */
1967         return 0;
1968 }
1969
1970 /**
1971  * Cancel all locks on a namespace (or a specific resource, if given)
1972  * that have 0 readers/writers.
1973  *
1974  * If flags & LCF_LOCAL, throw the locks away without trying
1975  * to notify the server. */
1976 int ldlm_cli_cancel_unused(struct ldlm_namespace *ns,
1977                            const struct ldlm_res_id *res_id,
1978                            ldlm_cancel_flags_t flags, void *opaque)
1979 {
1980         struct ldlm_cli_cancel_arg arg = {
1981                 .lc_flags       = flags,
1982                 .lc_opaque      = opaque,
1983         };
1984
1985         ENTRY;
1986
1987         if (ns == NULL)
1988                 RETURN(ELDLM_OK);
1989
1990         if (res_id != NULL) {
1991                 RETURN(ldlm_cli_cancel_unused_resource(ns, res_id, NULL,
1992                                                        LCK_MINMODE, flags,
1993                                                        opaque));
1994         } else {
1995                 cfs_hash_for_each_nolock(ns->ns_rs_hash,
1996                                          ldlm_cli_hash_cancel_unused, &arg);
1997                 RETURN(ELDLM_OK);
1998         }
1999 }
2000 EXPORT_SYMBOL(ldlm_cli_cancel_unused);
2001
2002 /* Lock iterators. */
2003
2004 int ldlm_resource_foreach(struct ldlm_resource *res, ldlm_iterator_t iter,
2005                           void *closure)
2006 {
2007         cfs_list_t *tmp, *next;
2008         struct ldlm_lock *lock;
2009         int rc = LDLM_ITER_CONTINUE;
2010
2011         ENTRY;
2012
2013         if (!res)
2014                 RETURN(LDLM_ITER_CONTINUE);
2015
2016         lock_res(res);
2017         cfs_list_for_each_safe(tmp, next, &res->lr_granted) {
2018                 lock = cfs_list_entry(tmp, struct ldlm_lock, l_res_link);
2019
2020                 if (iter(lock, closure) == LDLM_ITER_STOP)
2021                         GOTO(out, rc = LDLM_ITER_STOP);
2022         }
2023
2024         cfs_list_for_each_safe(tmp, next, &res->lr_converting) {
2025                 lock = cfs_list_entry(tmp, struct ldlm_lock, l_res_link);
2026
2027                 if (iter(lock, closure) == LDLM_ITER_STOP)
2028                         GOTO(out, rc = LDLM_ITER_STOP);
2029         }
2030
2031         cfs_list_for_each_safe(tmp, next, &res->lr_waiting) {
2032                 lock = cfs_list_entry(tmp, struct ldlm_lock, l_res_link);
2033
2034                 if (iter(lock, closure) == LDLM_ITER_STOP)
2035                         GOTO(out, rc = LDLM_ITER_STOP);
2036         }
2037  out:
2038         unlock_res(res);
2039         RETURN(rc);
2040 }
2041 EXPORT_SYMBOL(ldlm_resource_foreach);
2042
2043 struct iter_helper_data {
2044         ldlm_iterator_t iter;
2045         void *closure;
2046 };
2047
2048 static int ldlm_iter_helper(struct ldlm_lock *lock, void *closure)
2049 {
2050         struct iter_helper_data *helper = closure;
2051         return helper->iter(lock, helper->closure);
2052 }
2053
2054 static int ldlm_res_iter_helper(cfs_hash_t *hs, cfs_hash_bd_t *bd,
2055                                 cfs_hlist_node_t *hnode, void *arg)
2056
2057 {
2058         struct ldlm_resource *res = cfs_hash_object(hs, hnode);
2059
2060         return ldlm_resource_foreach(res, ldlm_iter_helper, arg) ==
2061                LDLM_ITER_STOP;
2062 }
2063
2064 void ldlm_namespace_foreach(struct ldlm_namespace *ns,
2065                             ldlm_iterator_t iter, void *closure)
2066
2067 {
2068         struct iter_helper_data helper = { iter: iter, closure: closure };
2069
2070         cfs_hash_for_each_nolock(ns->ns_rs_hash,
2071                                  ldlm_res_iter_helper, &helper);
2072
2073 }
2074 EXPORT_SYMBOL(ldlm_namespace_foreach);
2075
2076 /* non-blocking function to manipulate a lock whose cb_data is being put away.
2077  * return  0:  find no resource
2078  *       > 0:  must be LDLM_ITER_STOP/LDLM_ITER_CONTINUE.
2079  *       < 0:  errors
2080  */
2081 int ldlm_resource_iterate(struct ldlm_namespace *ns,
2082                           const struct ldlm_res_id *res_id,
2083                           ldlm_iterator_t iter, void *data)
2084 {
2085         struct ldlm_resource *res;
2086         int rc;
2087         ENTRY;
2088
2089         if (ns == NULL) {
2090                 CERROR("must pass in namespace\n");
2091                 LBUG();
2092         }
2093
2094         res = ldlm_resource_get(ns, NULL, res_id, 0, 0);
2095         if (res == NULL)
2096                 RETURN(0);
2097
2098         LDLM_RESOURCE_ADDREF(res);
2099         rc = ldlm_resource_foreach(res, iter, data);
2100         LDLM_RESOURCE_DELREF(res);
2101         ldlm_resource_putref(res);
2102         RETURN(rc);
2103 }
2104 EXPORT_SYMBOL(ldlm_resource_iterate);
2105
2106 /* Lock replay */
2107
2108 static int ldlm_chain_lock_for_replay(struct ldlm_lock *lock, void *closure)
2109 {
2110         cfs_list_t *list = closure;
2111
2112         /* we use l_pending_chain here, because it's unused on clients. */
2113         LASSERTF(cfs_list_empty(&lock->l_pending_chain),
2114                  "lock %p next %p prev %p\n",
2115                  lock, &lock->l_pending_chain.next,&lock->l_pending_chain.prev);
2116         /* bug 9573: don't replay locks left after eviction, or
2117          * bug 17614: locks being actively cancelled. Get a reference
2118          * on a lock so that it does not disapear under us (e.g. due to cancel)
2119          */
2120         if (!(lock->l_flags & (LDLM_FL_FAILED|LDLM_FL_CANCELING))) {
2121                 cfs_list_add(&lock->l_pending_chain, list);
2122                 LDLM_LOCK_GET(lock);
2123         }
2124
2125         return LDLM_ITER_CONTINUE;
2126 }
2127
2128 static int replay_lock_interpret(const struct lu_env *env,
2129                                  struct ptlrpc_request *req,
2130                                  struct ldlm_async_args *aa, int rc)
2131 {
2132         struct ldlm_lock     *lock;
2133         struct ldlm_reply    *reply;
2134         struct obd_export    *exp;
2135
2136         ENTRY;
2137         cfs_atomic_dec(&req->rq_import->imp_replay_inflight);
2138         if (rc != ELDLM_OK)
2139                 GOTO(out, rc);
2140
2141
2142         reply = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
2143         if (reply == NULL)
2144                 GOTO(out, rc = -EPROTO);
2145
2146         lock = ldlm_handle2lock(&aa->lock_handle);
2147         if (!lock) {
2148                 CERROR("received replay ack for unknown local cookie "LPX64
2149                        " remote cookie "LPX64 " from server %s id %s\n",
2150                        aa->lock_handle.cookie, reply->lock_handle.cookie,
2151                        req->rq_export->exp_client_uuid.uuid,
2152                        libcfs_id2str(req->rq_peer));
2153                 GOTO(out, rc = -ESTALE);
2154         }
2155
2156         /* Key change rehash lock in per-export hash with new key */
2157         exp = req->rq_export;
2158         if (exp && exp->exp_lock_hash) {
2159                 /* In the function below, .hs_keycmp resolves to
2160                  * ldlm_export_lock_keycmp() */
2161                 /* coverity[overrun-buffer-val] */
2162                 cfs_hash_rehash_key(exp->exp_lock_hash,
2163                                     &lock->l_remote_handle,
2164                                     &reply->lock_handle,
2165                                     &lock->l_exp_hash);
2166         } else {
2167                 lock->l_remote_handle = reply->lock_handle;
2168         }
2169
2170         LDLM_DEBUG(lock, "replayed lock:");
2171         ptlrpc_import_recovery_state_machine(req->rq_import);
2172         LDLM_LOCK_PUT(lock);
2173 out:
2174         if (rc != ELDLM_OK)
2175                 ptlrpc_connect_import(req->rq_import);
2176
2177         RETURN(rc);
2178 }
2179
2180 static int replay_one_lock(struct obd_import *imp, struct ldlm_lock *lock)
2181 {
2182         struct ptlrpc_request *req;
2183         struct ldlm_async_args *aa;
2184         struct ldlm_request   *body;
2185         int flags;
2186         ENTRY;
2187
2188
2189         /* Bug 11974: Do not replay a lock which is actively being canceled */
2190         if (lock->l_flags & LDLM_FL_CANCELING) {
2191                 LDLM_DEBUG(lock, "Not replaying canceled lock:");
2192                 RETURN(0);
2193         }
2194
2195         /* If this is reply-less callback lock, we cannot replay it, since
2196          * server might have long dropped it, but notification of that event was
2197          * lost by network. (and server granted conflicting lock already) */
2198         if (lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK) {
2199                 LDLM_DEBUG(lock, "Not replaying reply-less lock:");
2200                 ldlm_lock_cancel(lock);
2201                 RETURN(0);
2202         }
2203
2204         /*
2205          * If granted mode matches the requested mode, this lock is granted.
2206          *
2207          * If they differ, but we have a granted mode, then we were granted
2208          * one mode and now want another: ergo, converting.
2209          *
2210          * If we haven't been granted anything and are on a resource list,
2211          * then we're blocked/waiting.
2212          *
2213          * If we haven't been granted anything and we're NOT on a resource list,
2214          * then we haven't got a reply yet and don't have a known disposition.
2215          * This happens whenever a lock enqueue is the request that triggers
2216          * recovery.
2217          */
2218         if (lock->l_granted_mode == lock->l_req_mode)
2219                 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_GRANTED;
2220         else if (lock->l_granted_mode)
2221                 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_CONV;
2222         else if (!cfs_list_empty(&lock->l_res_link))
2223                 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_WAIT;
2224         else
2225                 flags = LDLM_FL_REPLAY;
2226
2227         req = ptlrpc_request_alloc_pack(imp, &RQF_LDLM_ENQUEUE,
2228                                         LUSTRE_DLM_VERSION, LDLM_ENQUEUE);
2229         if (req == NULL)
2230                 RETURN(-ENOMEM);
2231
2232         /* We're part of recovery, so don't wait for it. */
2233         req->rq_send_state = LUSTRE_IMP_REPLAY_LOCKS;
2234
2235         body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2236         ldlm_lock2desc(lock, &body->lock_desc);
2237         body->lock_flags = ldlm_flags_to_wire(flags);
2238
2239         ldlm_lock2handle(lock, &body->lock_handle[0]);
2240         if (lock->l_lvb_len > 0)
2241                 req_capsule_extend(&req->rq_pill, &RQF_LDLM_ENQUEUE_LVB);
2242         req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
2243                              lock->l_lvb_len);
2244         ptlrpc_request_set_replen(req);
2245         /* notify the server we've replayed all requests.
2246          * also, we mark the request to be put on a dedicated
2247          * queue to be processed after all request replayes.
2248          * bug 6063 */
2249         lustre_msg_set_flags(req->rq_reqmsg, MSG_REQ_REPLAY_DONE);
2250
2251         LDLM_DEBUG(lock, "replaying lock:");
2252
2253         cfs_atomic_inc(&req->rq_import->imp_replay_inflight);
2254         CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
2255         aa = ptlrpc_req_async_args(req);
2256         aa->lock_handle = body->lock_handle[0];
2257         req->rq_interpret_reply = (ptlrpc_interpterer_t)replay_lock_interpret;
2258         ptlrpcd_add_req(req, PDL_POLICY_LOCAL, -1);
2259
2260         RETURN(0);
2261 }
2262
2263 /**
2264  * Cancel as many unused locks as possible before replay. since we are
2265  * in recovery, we can't wait for any outstanding RPCs to send any RPC
2266  * to the server.
2267  *
2268  * Called only in recovery before replaying locks. there is no need to
2269  * replay locks that are unused. since the clients may hold thousands of
2270  * cached unused locks, dropping the unused locks can greatly reduce the
2271  * load on the servers at recovery time.
2272  */
2273 static void ldlm_cancel_unused_locks_for_replay(struct ldlm_namespace *ns)
2274 {
2275         int canceled;
2276         CFS_LIST_HEAD(cancels);
2277
2278         CDEBUG(D_DLMTRACE, "Dropping as many unused locks as possible before"
2279                            "replay for namespace %s (%d)\n",
2280                            ldlm_ns_name(ns), ns->ns_nr_unused);
2281
2282         /* We don't need to care whether or not LRU resize is enabled
2283          * because the LDLM_CANCEL_NO_WAIT policy doesn't use the
2284          * count parameter */
2285         canceled = ldlm_cancel_lru_local(ns, &cancels, ns->ns_nr_unused, 0,
2286                                          LCF_LOCAL, LDLM_CANCEL_NO_WAIT);
2287
2288         CDEBUG(D_DLMTRACE, "Canceled %d unused locks from namespace %s\n",
2289                            canceled, ldlm_ns_name(ns));
2290 }
2291
2292 int ldlm_replay_locks(struct obd_import *imp)
2293 {
2294         struct ldlm_namespace *ns = imp->imp_obd->obd_namespace;
2295         CFS_LIST_HEAD(list);
2296         struct ldlm_lock *lock, *next;
2297         int rc = 0;
2298
2299         ENTRY;
2300
2301         LASSERT(cfs_atomic_read(&imp->imp_replay_inflight) == 0);
2302
2303         /* don't replay locks if import failed recovery */
2304         if (imp->imp_vbr_failed)
2305                 RETURN(0);
2306
2307         /* ensure this doesn't fall to 0 before all have been queued */
2308         cfs_atomic_inc(&imp->imp_replay_inflight);
2309
2310         if (ldlm_cancel_unused_locks_before_replay)
2311                 ldlm_cancel_unused_locks_for_replay(ns);
2312
2313         ldlm_namespace_foreach(ns, ldlm_chain_lock_for_replay, &list);
2314
2315         cfs_list_for_each_entry_safe(lock, next, &list, l_pending_chain) {
2316                 cfs_list_del_init(&lock->l_pending_chain);
2317                 if (rc) {
2318                         LDLM_LOCK_RELEASE(lock);
2319                         continue; /* or try to do the rest? */
2320                 }
2321                 rc = replay_one_lock(imp, lock);
2322                 LDLM_LOCK_RELEASE(lock);
2323         }
2324
2325         cfs_atomic_dec(&imp->imp_replay_inflight);
2326
2327         RETURN(rc);
2328 }
2329 EXPORT_SYMBOL(ldlm_replay_locks);