Whamcloud - gitweb
LU-6325 ptlrpc: make ptlrpcd threads cpt-aware
[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, 2014, 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
64 #include <lustre_dlm.h>
65 #include <obd_class.h>
66 #include <obd.h>
67
68 #include "ldlm_internal.h"
69
70 unsigned int ldlm_enqueue_min = OBD_TIMEOUT_DEFAULT;
71 CFS_MODULE_PARM(ldlm_enqueue_min, "i", uint, 0644,
72                 "lock enqueue timeout minimum");
73
74 /* in client side, whether the cached locks will be canceled before replay */
75 unsigned int ldlm_cancel_unused_locks_before_replay = 1;
76
77 static void interrupted_completion_wait(void *data)
78 {
79 }
80
81 struct lock_wait_data {
82         struct ldlm_lock *lwd_lock;
83         __u32             lwd_conn_cnt;
84 };
85
86 struct ldlm_async_args {
87         struct lustre_handle lock_handle;
88 };
89
90 int ldlm_expired_completion_wait(void *data)
91 {
92         struct lock_wait_data *lwd = data;
93         struct ldlm_lock *lock = lwd->lwd_lock;
94         struct obd_import *imp;
95         struct obd_device *obd;
96
97         ENTRY;
98         if (lock->l_conn_export == NULL) {
99                 static cfs_time_t next_dump = 0, last_dump = 0;
100
101                 LCONSOLE_WARN("lock timed out (enqueued at "CFS_TIME_T", "
102                               CFS_DURATION_T"s ago)\n",
103                               lock->l_last_activity,
104                               cfs_time_sub(cfs_time_current_sec(),
105                                            lock->l_last_activity));
106                 LDLM_DEBUG(lock, "lock timed out (enqueued at "CFS_TIME_T", "
107                            CFS_DURATION_T"s ago); not entering recovery in "
108                            "server code, just going back to sleep",
109                            lock->l_last_activity,
110                            cfs_time_sub(cfs_time_current_sec(),
111                                         lock->l_last_activity));
112                 if (cfs_time_after(cfs_time_current(), next_dump)) {
113                         last_dump = next_dump;
114                         next_dump = cfs_time_shift(300);
115                         ldlm_namespace_dump(D_DLMTRACE,
116                                             ldlm_lock_to_ns(lock));
117                         if (last_dump == 0)
118                                 libcfs_debug_dumplog();
119                 }
120                 RETURN(0);
121         }
122
123         obd = lock->l_conn_export->exp_obd;
124         imp = obd->u.cli.cl_import;
125         ptlrpc_fail_import(imp, lwd->lwd_conn_cnt);
126         LDLM_ERROR(lock, "lock timed out (enqueued at "CFS_TIME_T", "
127                   CFS_DURATION_T"s ago), entering recovery for %s@%s",
128                   lock->l_last_activity,
129                   cfs_time_sub(cfs_time_current_sec(), lock->l_last_activity),
130                   obd2cli_tgt(obd), imp->imp_connection->c_remote_uuid.uuid);
131
132         RETURN(0);
133 }
134
135 /**
136  * Calculate the Completion timeout (covering enqueue, BL AST, data flush,
137  * lock cancel, and their replies). Used for lock completion timeout on the
138  * client side.
139  *
140  * \param[in] lock        lock which is waiting the completion callback
141  *
142  * \retval            timeout in seconds to wait for the server reply
143  */
144
145 /* We use the same basis for both server side and client side functions
146    from a single node. */
147 static unsigned int ldlm_cp_timeout(struct ldlm_lock *lock)
148 {
149         unsigned int timeout;
150
151         if (AT_OFF)
152                 return obd_timeout;
153
154         /* Wait a long time for enqueue - server may have to callback a
155          * lock from another client.  Server will evict the other client if it
156          * doesn't respond reasonably, and then give us the lock. */
157         timeout = at_get(ldlm_lock_to_ns_at(lock));
158         return max(3 * timeout, ldlm_enqueue_min);
159 }
160
161 /**
162  * Helper function for ldlm_completion_ast(), updating timings when lock is
163  * actually granted.
164  */
165 static int ldlm_completion_tail(struct ldlm_lock *lock, void *data)
166 {
167         long delay;
168         int  result = 0;
169
170         if (ldlm_is_destroyed(lock) || ldlm_is_failed(lock)) {
171                 LDLM_DEBUG(lock, "client-side enqueue: destroyed");
172                 result = -EIO;
173         } else if (data == NULL) {
174                 LDLM_DEBUG(lock, "client-side enqueue: granted");
175         } else {
176                 /* Take into AT only CP RPC, not immediately granted locks */
177                 delay = cfs_time_sub(cfs_time_current_sec(),
178                                      lock->l_last_activity);
179                 LDLM_DEBUG(lock, "client-side enqueue: granted after "
180                            CFS_DURATION_T"s", delay);
181
182                 /* Update our time estimate */
183                 at_measured(ldlm_lock_to_ns_at(lock), delay);
184         }
185         return result;
186 }
187
188 /**
189  * Implementation of ->l_completion_ast() for a client, that doesn't wait
190  * until lock is granted. Suitable for locks enqueued through ptlrpcd, of
191  * other threads that cannot block for long.
192  */
193 int ldlm_completion_ast_async(struct ldlm_lock *lock, __u64 flags, void *data)
194 {
195         ENTRY;
196
197         if (flags == LDLM_FL_WAIT_NOREPROC) {
198                 LDLM_DEBUG(lock, "client-side enqueue waiting on pending lock");
199                 RETURN(0);
200         }
201
202         if (!(flags & LDLM_FL_BLOCKED_MASK)) {
203                 wake_up(&lock->l_waitq);
204                 RETURN(ldlm_completion_tail(lock, data));
205         }
206
207         LDLM_DEBUG(lock, "client-side enqueue returned a blocked lock, "
208                    "going forward");
209         ldlm_reprocess_all(lock->l_resource);
210         RETURN(0);
211 }
212 EXPORT_SYMBOL(ldlm_completion_ast_async);
213
214 /**
215  * Generic LDLM "completion" AST. This is called in several cases:
216  *
217  *     - when a reply to an ENQUEUE RPC is received from the server
218  *       (ldlm_cli_enqueue_fini()). Lock might be granted or not granted at
219  *       this point (determined by flags);
220  *
221  *     - when LDLM_CP_CALLBACK RPC comes to client to notify it that lock has
222  *       been granted;
223  *
224  *     - when ldlm_lock_match(LDLM_FL_LVB_READY) is about to wait until lock
225  *       gets correct lvb;
226  *
227  *     - to force all locks when resource is destroyed (cleanup_resource());
228  *
229  *     - during lock conversion (not used currently).
230  *
231  * If lock is not granted in the first case, this function waits until second
232  * or penultimate cases happen in some other thread.
233  *
234  */
235 int ldlm_completion_ast(struct ldlm_lock *lock, __u64 flags, void *data)
236 {
237         /* XXX ALLOCATE - 160 bytes */
238         struct lock_wait_data lwd;
239         struct obd_device *obd;
240         struct obd_import *imp = NULL;
241         struct l_wait_info lwi;
242         __u32 timeout;
243         int rc = 0;
244         ENTRY;
245
246         if (flags == LDLM_FL_WAIT_NOREPROC) {
247                 LDLM_DEBUG(lock, "client-side enqueue waiting on pending lock");
248                 goto noreproc;
249         }
250
251         if (!(flags & LDLM_FL_BLOCKED_MASK)) {
252                 wake_up(&lock->l_waitq);
253                 RETURN(0);
254         }
255
256         LDLM_DEBUG(lock, "client-side enqueue returned a blocked lock, "
257                    "sleeping");
258
259 noreproc:
260
261         obd = class_exp2obd(lock->l_conn_export);
262
263         /* if this is a local lock, then there is no import */
264         if (obd != NULL) {
265                 imp = obd->u.cli.cl_import;
266         }
267
268         timeout = ldlm_cp_timeout(lock);
269
270         lwd.lwd_lock = lock;
271         lock->l_last_activity = cfs_time_current_sec();
272
273         if (ldlm_is_no_timeout(lock)) {
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                 ldlm_set_fail_loc(lock);
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, data));
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         ldlm_set_cbpending(lock);
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  * Implements ldlm_lock::l_glimpse_ast for extent locks acquired on the server.
385  *
386  * Returning -ELDLM_NO_LOCK_DATA actually works, but the reason for that is
387  * rather subtle: with OST-side locking, it may so happen that _all_ extent
388  * locks are held by the OST. If client wants to obtain the current file size
389  * it calls ll_glimpse_size(), and (as all locks are held only on the server),
390  * this dummy glimpse callback fires and does nothing. The client still
391  * receives the correct file size due to the following fragment of code in
392  * ldlm_cb_interpret():
393  *
394  *      if (rc == -ELDLM_NO_LOCK_DATA) {
395  *              LDLM_DEBUG(lock, "lost race - client has a lock but no"
396  *                         "inode");
397  *              ldlm_res_lvbo_update(lock->l_resource, NULL, 1);
398  *      }
399  *
400  * That is, after the glimpse returns this error, ofd_lvbo_update() is called
401  * and returns the updated file attributes from the inode to the client.
402  *
403  * See also comment in ofd_intent_policy() on why servers must set a non-NULL
404  * l_glimpse_ast when grabbing DLM locks.  Otherwise, the server will assume
405  * that the object is in the process of being destroyed.
406  *
407  * \param[in] lock      DLM lock being glimpsed, unused
408  * \param[in] reqp      pointer to ptlrpc_request, unused
409  *
410  * \retval              -ELDLM_NO_LOCK_DATA to get attributes from disk object
411  */
412 int ldlm_glimpse_ast(struct ldlm_lock *lock, void *reqp)
413 {
414         return -ELDLM_NO_LOCK_DATA;
415 }
416
417 /**
418  * Enqueue a local lock (typically on a server).
419  */
420 int ldlm_cli_enqueue_local(struct ldlm_namespace *ns,
421                            const struct ldlm_res_id *res_id,
422                            ldlm_type_t type, ldlm_policy_data_t *policy,
423                            ldlm_mode_t mode, __u64 *flags,
424                            ldlm_blocking_callback blocking,
425                            ldlm_completion_callback completion,
426                            ldlm_glimpse_callback glimpse,
427                            void *data, __u32 lvb_len, enum lvb_type lvb_type,
428                            const __u64 *client_cookie,
429                            struct lustre_handle *lockh)
430 {
431         struct ldlm_lock *lock;
432         int err;
433         const struct ldlm_callback_suite cbs = { .lcs_completion = completion,
434                                                  .lcs_blocking   = blocking,
435                                                  .lcs_glimpse    = glimpse,
436         };
437         ENTRY;
438
439         LASSERT(!(*flags & LDLM_FL_REPLAY));
440         if (unlikely(ns_is_client(ns))) {
441                 CERROR("Trying to enqueue local lock in a shadow namespace\n");
442                 LBUG();
443         }
444
445         lock = ldlm_lock_create(ns, res_id, type, mode, &cbs, data, lvb_len,
446                                 lvb_type);
447         if (IS_ERR(lock))
448                 GOTO(out_nolock, err = PTR_ERR(lock));
449
450         err = ldlm_lvbo_init(lock->l_resource);
451         if (err < 0) {
452                 LDLM_ERROR(lock, "delayed lvb init failed (rc %d)", err);
453                 GOTO(out, err);
454         }
455
456         ldlm_lock2handle(lock, lockh);
457
458         /* NB: we don't have any lock now (lock_res_and_lock)
459          * because it's a new lock */
460         ldlm_lock_addref_internal_nolock(lock, mode);
461         ldlm_set_local(lock);
462         if (*flags & LDLM_FL_ATOMIC_CB)
463                 ldlm_set_atomic_cb(lock);
464
465         if (policy != NULL)
466                 lock->l_policy_data = *policy;
467         if (client_cookie != NULL)
468                 lock->l_client_cookie = *client_cookie;
469         if (type == LDLM_EXTENT) {
470                 /* extent lock without policy is a bug */
471                 if (policy == NULL)
472                         LBUG();
473
474                 lock->l_req_extent = policy->l_extent;
475         }
476
477         err = ldlm_lock_enqueue(ns, &lock, policy, flags);
478         if (unlikely(err != ELDLM_OK))
479                 GOTO(out, err);
480
481         if (policy != NULL)
482                 *policy = lock->l_policy_data;
483
484         if (lock->l_completion_ast)
485                 lock->l_completion_ast(lock, *flags, NULL);
486
487         LDLM_DEBUG(lock, "client-side local enqueue handler, new lock created");
488         EXIT;
489  out:
490         LDLM_LOCK_RELEASE(lock);
491  out_nolock:
492         return err;
493 }
494 EXPORT_SYMBOL(ldlm_cli_enqueue_local);
495
496 static void failed_lock_cleanup(struct ldlm_namespace *ns,
497                                 struct ldlm_lock *lock, int mode)
498 {
499         int need_cancel = 0;
500
501         /* Set a flag to prevent us from sending a CANCEL (bug 407) */
502         lock_res_and_lock(lock);
503         /* Check that lock is not granted or failed, we might race. */
504         if ((lock->l_req_mode != lock->l_granted_mode) &&
505             !ldlm_is_failed(lock)) {
506                 /* Make sure that this lock will not be found by raced
507                  * bl_ast and -EINVAL reply is sent to server anyways.
508                  * b=17645*/
509                 lock->l_flags |= LDLM_FL_LOCAL_ONLY | LDLM_FL_FAILED |
510                                  LDLM_FL_ATOMIC_CB | LDLM_FL_CBPENDING;
511                 need_cancel = 1;
512         }
513         unlock_res_and_lock(lock);
514
515         if (need_cancel)
516                 LDLM_DEBUG(lock,
517                            "setting FL_LOCAL_ONLY | LDLM_FL_FAILED | "
518                            "LDLM_FL_ATOMIC_CB | LDLM_FL_CBPENDING");
519         else
520                 LDLM_DEBUG(lock, "lock was granted or failed in race");
521
522         /* XXX - HACK because we shouldn't call ldlm_lock_destroy()
523          *       from llite/file.c/ll_file_flock(). */
524         /* This code makes for the fact that we do not have blocking handler on
525          * a client for flock locks. As such this is the place where we must
526          * completely kill failed locks. (interrupted and those that
527          * were waiting to be granted when server evicted us. */
528         if (lock->l_resource->lr_type == LDLM_FLOCK) {
529                 lock_res_and_lock(lock);
530                 if (!ldlm_is_destroyed(lock)) {
531                         ldlm_resource_unlink_lock(lock);
532                         ldlm_lock_decref_internal_nolock(lock, mode);
533                         ldlm_lock_destroy_nolock(lock);
534                 }
535                 unlock_res_and_lock(lock);
536         } else {
537                 ldlm_lock_decref_internal(lock, mode);
538         }
539 }
540
541 /**
542  * Finishing portion of client lock enqueue code.
543  *
544  * Called after receiving reply from server.
545  */
546 int ldlm_cli_enqueue_fini(struct obd_export *exp, struct ptlrpc_request *req,
547                           ldlm_type_t type, __u8 with_policy, ldlm_mode_t mode,
548                           __u64 *flags, void *lvb, __u32 lvb_len,
549                           struct lustre_handle *lockh,int rc)
550 {
551         struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
552         int is_replay = *flags & LDLM_FL_REPLAY;
553         struct ldlm_lock *lock;
554         struct ldlm_reply *reply;
555         int cleanup_phase = 1;
556         ENTRY;
557
558         lock = ldlm_handle2lock(lockh);
559         /* ldlm_cli_enqueue is holding a reference on this lock. */
560         if (!lock) {
561                 LASSERT(type == LDLM_FLOCK);
562                 RETURN(-ENOLCK);
563         }
564
565         LASSERTF(ergo(lvb_len != 0, lvb_len == lock->l_lvb_len),
566                  "lvb_len = %d, l_lvb_len = %d\n", lvb_len, lock->l_lvb_len);
567
568         if (rc != ELDLM_OK) {
569                 LASSERT(!is_replay);
570                 LDLM_DEBUG(lock, "client-side enqueue END (%s)",
571                            rc == ELDLM_LOCK_ABORTED ? "ABORTED" : "FAILED");
572
573                 if (rc != ELDLM_LOCK_ABORTED)
574                         GOTO(cleanup, rc);
575         }
576
577         /* Before we return, swab the reply */
578         reply = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
579         if (reply == NULL)
580                 GOTO(cleanup, rc = -EPROTO);
581
582         if (lvb_len > 0) {
583                 int size = 0;
584
585                 size = req_capsule_get_size(&req->rq_pill, &RMF_DLM_LVB,
586                                             RCL_SERVER);
587                 if (size < 0) {
588                         LDLM_ERROR(lock, "Fail to get lvb_len, rc = %d", size);
589                         GOTO(cleanup, rc = size);
590                 } else if (unlikely(size > lvb_len)) {
591                         LDLM_ERROR(lock, "Replied LVB is larger than "
592                                    "expectation, expected = %d, replied = %d",
593                                    lvb_len, size);
594                         GOTO(cleanup, rc = -EINVAL);
595                 }
596                 lvb_len = size;
597         }
598
599         if (rc == ELDLM_LOCK_ABORTED) {
600                 if (lvb_len > 0 && lvb != NULL)
601                         rc = ldlm_fill_lvb(lock, &req->rq_pill, RCL_SERVER,
602                                            lvb, lvb_len);
603                 GOTO(cleanup, rc = rc ? : ELDLM_LOCK_ABORTED);
604         }
605
606         /* lock enqueued on the server */
607         cleanup_phase = 0;
608
609         lock_res_and_lock(lock);
610         /* Key change rehash lock in per-export hash with new key */
611         if (exp->exp_lock_hash) {
612                 /* In the function below, .hs_keycmp resolves to
613                  * ldlm_export_lock_keycmp() */
614                 /* coverity[overrun-buffer-val] */
615                 cfs_hash_rehash_key(exp->exp_lock_hash,
616                                     &lock->l_remote_handle,
617                                     &reply->lock_handle,
618                                     &lock->l_exp_hash);
619         } else {
620                 lock->l_remote_handle = reply->lock_handle;
621         }
622
623         *flags = ldlm_flags_from_wire(reply->lock_flags);
624         lock->l_flags |= ldlm_flags_from_wire(reply->lock_flags &
625                                               LDLM_FL_INHERIT_MASK);
626         unlock_res_and_lock(lock);
627
628         CDEBUG(D_INFO, "local: %p, remote cookie: "LPX64", flags: "LPX64"\n",
629                lock, reply->lock_handle.cookie, *flags);
630
631         /* If enqueue returned a blocked lock but the completion handler has
632          * already run, then it fixed up the resource and we don't need to do it
633          * again. */
634         if ((*flags) & LDLM_FL_LOCK_CHANGED) {
635                 int newmode = reply->lock_desc.l_req_mode;
636                 LASSERT(!is_replay);
637                 if (newmode && newmode != lock->l_req_mode) {
638                         LDLM_DEBUG(lock, "server returned different mode %s",
639                                    ldlm_lockname[newmode]);
640                         lock->l_req_mode = newmode;
641                 }
642
643                 if (!ldlm_res_eq(&reply->lock_desc.l_resource.lr_name,
644                                  &lock->l_resource->lr_name)) {
645                         CDEBUG(D_INFO, "remote intent success, locking "DLDLMRES
646                                        " instead of "DLDLMRES"\n",
647                                PLDLMRES(&reply->lock_desc.l_resource),
648                                PLDLMRES(lock->l_resource));
649
650                         rc = ldlm_lock_change_resource(ns, lock,
651                                         &reply->lock_desc.l_resource.lr_name);
652                         if (rc || lock->l_resource == NULL)
653                                 GOTO(cleanup, rc = -ENOMEM);
654                         LDLM_DEBUG(lock, "client-side enqueue, new resource");
655                 }
656                 if (with_policy)
657                         if (!(type == LDLM_IBITS &&
658                               !(exp_connect_flags(exp) & OBD_CONNECT_IBITS)))
659                                 /* We assume lock type cannot change on server*/
660                                 ldlm_convert_policy_to_local(exp,
661                                                 lock->l_resource->lr_type,
662                                                 &reply->lock_desc.l_policy_data,
663                                                 &lock->l_policy_data);
664                 if (type != LDLM_PLAIN)
665                         LDLM_DEBUG(lock,"client-side enqueue, new policy data");
666         }
667
668         if ((*flags) & LDLM_FL_AST_SENT) {
669                 lock_res_and_lock(lock);
670                 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_BL_AST;
671                 unlock_res_and_lock(lock);
672                 LDLM_DEBUG(lock, "enqueue reply includes blocking AST");
673         }
674
675         /* If the lock has already been granted by a completion AST, don't
676          * clobber the LVB with an older one. */
677         if (lvb_len > 0) {
678                 /* We must lock or a racing completion might update lvb without
679                  * letting us know and we'll clobber the correct value.
680                  * Cannot unlock after the check either, a that still leaves
681                  * a tiny window for completion to get in */
682                 lock_res_and_lock(lock);
683                 if (lock->l_req_mode != lock->l_granted_mode)
684                         rc = ldlm_fill_lvb(lock, &req->rq_pill, RCL_SERVER,
685                                            lock->l_lvb_data, lvb_len);
686                 unlock_res_and_lock(lock);
687                 if (rc < 0) {
688                         cleanup_phase = 1;
689                         GOTO(cleanup, rc);
690                 }
691         }
692
693         if (!is_replay) {
694                 rc = ldlm_lock_enqueue(ns, &lock, NULL, flags);
695                 if (lock->l_completion_ast != NULL) {
696                         int err = lock->l_completion_ast(lock, *flags, NULL);
697                         if (!rc)
698                                 rc = err;
699                         if (rc)
700                                 cleanup_phase = 1;
701                 }
702         }
703
704         if (lvb_len > 0 && lvb != NULL) {
705                 /* Copy the LVB here, and not earlier, because the completion
706                  * AST (if any) can override what we got in the reply */
707                 memcpy(lvb, lock->l_lvb_data, lvb_len);
708         }
709
710         LDLM_DEBUG(lock, "client-side enqueue END");
711         EXIT;
712 cleanup:
713         if (cleanup_phase == 1 && rc)
714                 failed_lock_cleanup(ns, lock, mode);
715         /* Put lock 2 times, the second reference is held by ldlm_cli_enqueue */
716         LDLM_LOCK_PUT(lock);
717         LDLM_LOCK_RELEASE(lock);
718         return rc;
719 }
720 EXPORT_SYMBOL(ldlm_cli_enqueue_fini);
721
722 /**
723  * Estimate number of lock handles that would fit into request of given
724  * size.  PAGE_SIZE-512 is to allow TCP/IP and LNET headers to fit into
725  * a single page on the send/receive side. XXX: 512 should be changed to
726  * more adequate value.
727  */
728 static inline int ldlm_req_handles_avail(int req_size, int off)
729 {
730         int avail;
731
732         avail = min_t(int, LDLM_MAXREQSIZE, PAGE_CACHE_SIZE - 512) - req_size;
733         if (likely(avail >= 0))
734                 avail /= (int)sizeof(struct lustre_handle);
735         else
736                 avail = 0;
737         avail += LDLM_LOCKREQ_HANDLES - off;
738
739         return avail;
740 }
741
742 static inline int ldlm_capsule_handles_avail(struct req_capsule *pill,
743                                              enum req_location loc,
744                                              int off)
745 {
746         __u32 size = req_capsule_msg_size(pill, loc);
747         return ldlm_req_handles_avail(size, off);
748 }
749
750 static inline int ldlm_format_handles_avail(struct obd_import *imp,
751                                             const struct req_format *fmt,
752                                             enum req_location loc, int off)
753 {
754         __u32 size = req_capsule_fmt_size(imp->imp_msg_magic, fmt, loc);
755         return ldlm_req_handles_avail(size, off);
756 }
757
758 /**
759  * Cancel LRU locks and pack them into the enqueue request. Pack there the given
760  * \a count locks in \a cancels.
761  *
762  * This is to be called by functions preparing their own requests that
763  * might contain lists of locks to cancel in addition to actual operation
764  * that needs to be performed.
765  */
766 int ldlm_prep_elc_req(struct obd_export *exp, struct ptlrpc_request *req,
767                       int version, int opc, int canceloff,
768                       struct list_head *cancels, int count)
769         {
770         struct ldlm_namespace   *ns = exp->exp_obd->obd_namespace;
771         struct req_capsule      *pill = &req->rq_pill;
772         struct ldlm_request     *dlm = NULL;
773         struct list_head        head = LIST_HEAD_INIT(head);
774         int flags, avail, to_free, pack = 0;
775         int rc;
776         ENTRY;
777
778         if (cancels == NULL)
779                 cancels = &head;
780         if (ns_connect_cancelset(ns)) {
781                 /* Estimate the amount of available space in the request. */
782                 req_capsule_filled_sizes(pill, RCL_CLIENT);
783                 avail = ldlm_capsule_handles_avail(pill, RCL_CLIENT, canceloff);
784
785                 flags = ns_connect_lru_resize(ns) ?
786                         LDLM_CANCEL_LRUR_NO_WAIT : LDLM_CANCEL_AGED;
787                 to_free = !ns_connect_lru_resize(ns) &&
788                           opc == LDLM_ENQUEUE ? 1 : 0;
789
790                 /* Cancel LRU locks here _only_ if the server supports
791                  * EARLY_CANCEL. Otherwise we have to send extra CANCEL
792                  * RPC, which will make us slower. */
793                 if (avail > count)
794                         count += ldlm_cancel_lru_local(ns, cancels, to_free,
795                                                        avail - count, 0, flags);
796                 if (avail > count)
797                         pack = count;
798                 else
799                         pack = avail;
800                 req_capsule_set_size(pill, &RMF_DLM_REQ, RCL_CLIENT,
801                                      ldlm_request_bufsize(pack, opc));
802         }
803
804         rc = ptlrpc_request_pack(req, version, opc);
805         if (rc) {
806                 ldlm_lock_list_put(cancels, l_bl_ast, count);
807                 RETURN(rc);
808         }
809
810         if (ns_connect_cancelset(ns)) {
811                 if (canceloff) {
812                         dlm = req_capsule_client_get(pill, &RMF_DLM_REQ);
813                         LASSERT(dlm);
814                         /* Skip first lock handler in ldlm_request_pack(),
815                          * this method will increment @lock_count according
816                          * to the lock handle amount actually written to
817                          * the buffer. */
818                         dlm->lock_count = canceloff;
819                 }
820                 /* Pack into the request @pack lock handles. */
821                 ldlm_cli_cancel_list(cancels, pack, req, 0);
822                 /* Prepare and send separate cancel RPC for others. */
823                 ldlm_cli_cancel_list(cancels, count - pack, NULL, 0);
824         } else {
825                 ldlm_lock_list_put(cancels, l_bl_ast, count);
826         }
827         RETURN(0);
828 }
829 EXPORT_SYMBOL(ldlm_prep_elc_req);
830
831 int ldlm_prep_enqueue_req(struct obd_export *exp, struct ptlrpc_request *req,
832                           struct list_head *cancels, int count)
833 {
834         return ldlm_prep_elc_req(exp, req, LUSTRE_DLM_VERSION, LDLM_ENQUEUE,
835                                  LDLM_ENQUEUE_CANCEL_OFF, cancels, count);
836 }
837 EXPORT_SYMBOL(ldlm_prep_enqueue_req);
838
839 struct ptlrpc_request *ldlm_enqueue_pack(struct obd_export *exp, int lvb_len)
840 {
841         struct ptlrpc_request *req;
842         int rc;
843         ENTRY;
844
845         req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_LDLM_ENQUEUE);
846         if (req == NULL)
847                 RETURN(ERR_PTR(-ENOMEM));
848
849         rc = ldlm_prep_enqueue_req(exp, req, NULL, 0);
850         if (rc) {
851                 ptlrpc_request_free(req);
852                 RETURN(ERR_PTR(rc));
853         }
854
855         req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER, lvb_len);
856         ptlrpc_request_set_replen(req);
857         RETURN(req);
858 }
859 EXPORT_SYMBOL(ldlm_enqueue_pack);
860
861 /**
862  * Client-side lock enqueue.
863  *
864  * If a request has some specific initialisation it is passed in \a reqp,
865  * otherwise it is created in ldlm_cli_enqueue.
866  *
867  * Supports sync and async requests, pass \a async flag accordingly. If a
868  * request was created in ldlm_cli_enqueue and it is the async request,
869  * pass it to the caller in \a reqp.
870  */
871 int ldlm_cli_enqueue(struct obd_export *exp, struct ptlrpc_request **reqp,
872                      struct ldlm_enqueue_info *einfo,
873                      const struct ldlm_res_id *res_id,
874                      ldlm_policy_data_t const *policy, __u64 *flags,
875                      void *lvb, __u32 lvb_len, enum lvb_type lvb_type,
876                      struct lustre_handle *lockh, int async)
877 {
878         struct ldlm_namespace *ns;
879         struct ldlm_lock      *lock;
880         struct ldlm_request   *body;
881         int                    is_replay = *flags & LDLM_FL_REPLAY;
882         int                    req_passed_in = 1;
883         int                    rc, err;
884         struct ptlrpc_request *req;
885         ENTRY;
886
887         LASSERT(exp != NULL);
888
889         ns = exp->exp_obd->obd_namespace;
890
891         /* If we're replaying this lock, just check some invariants.
892          * If we're creating a new lock, get everything all setup nice. */
893         if (is_replay) {
894                 lock = ldlm_handle2lock_long(lockh, 0);
895                 LASSERT(lock != NULL);
896                 LDLM_DEBUG(lock, "client-side enqueue START");
897                 LASSERT(exp == lock->l_conn_export);
898         } else {
899                 const struct ldlm_callback_suite cbs = {
900                         .lcs_completion = einfo->ei_cb_cp,
901                         .lcs_blocking   = einfo->ei_cb_bl,
902                         .lcs_glimpse    = einfo->ei_cb_gl
903                 };
904                 lock = ldlm_lock_create(ns, res_id, einfo->ei_type,
905                                         einfo->ei_mode, &cbs, einfo->ei_cbdata,
906                                         lvb_len, lvb_type);
907                 if (IS_ERR(lock))
908                         RETURN(PTR_ERR(lock));
909                 /* for the local lock, add the reference */
910                 ldlm_lock_addref_internal(lock, einfo->ei_mode);
911                 ldlm_lock2handle(lock, lockh);
912                 if (policy != NULL)
913                         lock->l_policy_data = *policy;
914
915                 if (einfo->ei_type == LDLM_EXTENT) {
916                         /* extent lock without policy is a bug */
917                         if (policy == NULL)
918                                 LBUG();
919
920                         lock->l_req_extent = policy->l_extent;
921                 }
922                 LDLM_DEBUG(lock, "client-side enqueue START, flags "LPX64"\n",
923                            *flags);
924         }
925
926         lock->l_conn_export = exp;
927         lock->l_export = NULL;
928         lock->l_blocking_ast = einfo->ei_cb_bl;
929         lock->l_flags |= (*flags & (LDLM_FL_NO_LRU | LDLM_FL_EXCL));
930         lock->l_last_activity = cfs_time_current_sec();
931
932         /* lock not sent to server yet */
933
934         if (reqp == NULL || *reqp == NULL) {
935                 req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
936                                                 &RQF_LDLM_ENQUEUE,
937                                                 LUSTRE_DLM_VERSION,
938                                                 LDLM_ENQUEUE);
939                 if (req == NULL) {
940                         failed_lock_cleanup(ns, lock, einfo->ei_mode);
941                         LDLM_LOCK_RELEASE(lock);
942                         RETURN(-ENOMEM);
943                 }
944                 req_passed_in = 0;
945                 if (reqp)
946                         *reqp = req;
947         } else {
948                 int len;
949
950                 req = *reqp;
951                 len = req_capsule_get_size(&req->rq_pill, &RMF_DLM_REQ,
952                                            RCL_CLIENT);
953                 LASSERTF(len >= sizeof(*body), "buflen[%d] = %d, not %d\n",
954                          DLM_LOCKREQ_OFF, len, (int)sizeof(*body));
955         }
956
957         /* Dump lock data into the request buffer */
958         body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
959         ldlm_lock2desc(lock, &body->lock_desc);
960         body->lock_flags = ldlm_flags_to_wire(*flags);
961         body->lock_handle[0] = *lockh;
962
963         /* Continue as normal. */
964         if (!req_passed_in) {
965                 if (lvb_len > 0)
966                         req_capsule_extend(&req->rq_pill,
967                                            &RQF_LDLM_ENQUEUE_LVB);
968                 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
969                                      lvb_len);
970                 ptlrpc_request_set_replen(req);
971         }
972
973         if (async) {
974                 LASSERT(reqp != NULL);
975                 RETURN(0);
976         }
977
978         LDLM_DEBUG(lock, "sending request");
979
980         rc = ptlrpc_queue_wait(req);
981
982         err = ldlm_cli_enqueue_fini(exp, req, einfo->ei_type, policy ? 1 : 0,
983                                     einfo->ei_mode, flags, lvb, lvb_len,
984                                     lockh, rc);
985
986         /* If ldlm_cli_enqueue_fini did not find the lock, we need to free
987          * one reference that we took */
988         if (err == -ENOLCK)
989                 LDLM_LOCK_RELEASE(lock);
990         else
991                 rc = err;
992
993         if (!req_passed_in && req != NULL) {
994                 ptlrpc_req_finished(req);
995                 if (reqp)
996                         *reqp = NULL;
997         }
998
999         RETURN(rc);
1000 }
1001 EXPORT_SYMBOL(ldlm_cli_enqueue);
1002
1003 static int ldlm_cli_convert_local(struct ldlm_lock *lock, int new_mode,
1004                                   __u32 *flags)
1005 {
1006         struct ldlm_resource *res;
1007         int rc;
1008         ENTRY;
1009         if (ns_is_client(ldlm_lock_to_ns(lock))) {
1010                 CERROR("Trying to cancel local lock\n");
1011                 LBUG();
1012         }
1013         LDLM_DEBUG(lock, "client-side local convert");
1014
1015         res = ldlm_lock_convert(lock, new_mode, flags);
1016         if (res) {
1017                 ldlm_reprocess_all(res);
1018                 rc = 0;
1019         } else {
1020                 rc = LUSTRE_EDEADLK;
1021         }
1022         LDLM_DEBUG(lock, "client-side local convert handler END");
1023         LDLM_LOCK_PUT(lock);
1024         RETURN(rc);
1025 }
1026
1027 /* FIXME: one of ldlm_cli_convert or the server side should reject attempted
1028  * conversion of locks which are on the waiting or converting queue */
1029 /* Caller of this code is supposed to take care of lock readers/writers
1030    accounting */
1031 int ldlm_cli_convert(struct lustre_handle *lockh, int new_mode, __u32 *flags)
1032 {
1033         struct ldlm_request   *body;
1034         struct ldlm_reply     *reply;
1035         struct ldlm_lock      *lock;
1036         struct ldlm_resource  *res;
1037         struct ptlrpc_request *req;
1038         int                    rc;
1039         ENTRY;
1040
1041         lock = ldlm_handle2lock(lockh);
1042         if (!lock) {
1043                 LBUG();
1044                 RETURN(-EINVAL);
1045         }
1046         *flags = 0;
1047
1048         if (lock->l_conn_export == NULL)
1049                 RETURN(ldlm_cli_convert_local(lock, new_mode, flags));
1050
1051         LDLM_DEBUG(lock, "client-side convert");
1052
1053         req = ptlrpc_request_alloc_pack(class_exp2cliimp(lock->l_conn_export),
1054                                         &RQF_LDLM_CONVERT, LUSTRE_DLM_VERSION,
1055                                         LDLM_CONVERT);
1056         if (req == NULL) {
1057                 LDLM_LOCK_PUT(lock);
1058                 RETURN(-ENOMEM);
1059         }
1060
1061         body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1062         body->lock_handle[0] = lock->l_remote_handle;
1063
1064         body->lock_desc.l_req_mode = new_mode;
1065         body->lock_flags = ldlm_flags_to_wire(*flags);
1066
1067
1068         ptlrpc_request_set_replen(req);
1069         rc = ptlrpc_queue_wait(req);
1070         if (rc != ELDLM_OK)
1071                 GOTO(out, rc);
1072
1073         reply = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
1074         if (reply == NULL)
1075                 GOTO(out, rc = -EPROTO);
1076
1077         if (req->rq_status)
1078                 GOTO(out, rc = req->rq_status);
1079
1080         res = ldlm_lock_convert(lock, new_mode, &reply->lock_flags);
1081         if (res != NULL) {
1082                 ldlm_reprocess_all(res);
1083                 /* Go to sleep until the lock is granted. */
1084                 /* FIXME: or cancelled. */
1085                 if (lock->l_completion_ast) {
1086                         rc = lock->l_completion_ast(lock, LDLM_FL_WAIT_NOREPROC,
1087                                                     NULL);
1088                         if (rc)
1089                                 GOTO(out, rc);
1090                 }
1091         } else {
1092                 rc = LUSTRE_EDEADLK;
1093         }
1094         EXIT;
1095  out:
1096         LDLM_LOCK_PUT(lock);
1097         ptlrpc_req_finished(req);
1098         return rc;
1099 }
1100
1101 /**
1102  * Cancel locks locally.
1103  * Returns:
1104  * \retval LDLM_FL_LOCAL_ONLY if there is no need for a CANCEL RPC to the server
1105  * \retval LDLM_FL_CANCELING otherwise;
1106  * \retval LDLM_FL_BL_AST if there is a need for a separate CANCEL RPC.
1107  */
1108 static __u64 ldlm_cli_cancel_local(struct ldlm_lock *lock)
1109 {
1110         __u64 rc = LDLM_FL_LOCAL_ONLY;
1111         ENTRY;
1112
1113         if (lock->l_conn_export) {
1114                 bool local_only;
1115
1116                 LDLM_DEBUG(lock, "client-side cancel");
1117                 /* Set this flag to prevent others from getting new references*/
1118                 lock_res_and_lock(lock);
1119                 ldlm_set_cbpending(lock);
1120                 local_only = !!(lock->l_flags &
1121                                 (LDLM_FL_LOCAL_ONLY|LDLM_FL_CANCEL_ON_BLOCK));
1122                 ldlm_cancel_callback(lock);
1123                 rc = (ldlm_is_bl_ast(lock)) ?
1124                         LDLM_FL_BL_AST : LDLM_FL_CANCELING;
1125                 unlock_res_and_lock(lock);
1126
1127                 if (local_only) {
1128                         CDEBUG(D_DLMTRACE, "not sending request (at caller's "
1129                                "instruction)\n");
1130                         rc = LDLM_FL_LOCAL_ONLY;
1131                 }
1132                 ldlm_lock_cancel(lock);
1133         } else {
1134                 if (ns_is_client(ldlm_lock_to_ns(lock))) {
1135                         LDLM_ERROR(lock, "Trying to cancel local lock");
1136                         LBUG();
1137                 }
1138                 LDLM_DEBUG(lock, "server-side local cancel");
1139                 ldlm_lock_cancel(lock);
1140                 ldlm_reprocess_all(lock->l_resource);
1141         }
1142
1143         RETURN(rc);
1144 }
1145
1146 /**
1147  * Pack \a count locks in \a head into ldlm_request buffer of request \a req.
1148  */
1149 static void ldlm_cancel_pack(struct ptlrpc_request *req,
1150                              struct list_head *head, int count)
1151 {
1152         struct ldlm_request *dlm;
1153         struct ldlm_lock *lock;
1154         int max, packed = 0;
1155         ENTRY;
1156
1157         dlm = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1158         LASSERT(dlm != NULL);
1159
1160         /* Check the room in the request buffer. */
1161         max = req_capsule_get_size(&req->rq_pill, &RMF_DLM_REQ, RCL_CLIENT) -
1162                 sizeof(struct ldlm_request);
1163         max /= sizeof(struct lustre_handle);
1164         max += LDLM_LOCKREQ_HANDLES;
1165         LASSERT(max >= dlm->lock_count + count);
1166
1167         /* XXX: it would be better to pack lock handles grouped by resource.
1168          * so that the server cancel would call filter_lvbo_update() less
1169          * frequently. */
1170         list_for_each_entry(lock, head, l_bl_ast) {
1171                 if (!count--)
1172                         break;
1173                 LASSERT(lock->l_conn_export);
1174                 /* Pack the lock handle to the given request buffer. */
1175                 LDLM_DEBUG(lock, "packing");
1176                 dlm->lock_handle[dlm->lock_count++] = lock->l_remote_handle;
1177                 packed++;
1178         }
1179         CDEBUG(D_DLMTRACE, "%d locks packed\n", packed);
1180         EXIT;
1181 }
1182
1183 /**
1184  * Prepare and send a batched cancel RPC. It will include \a count lock
1185  * handles of locks given in \a cancels list. */
1186 int ldlm_cli_cancel_req(struct obd_export *exp, struct list_head *cancels,
1187                         int count, ldlm_cancel_flags_t flags)
1188 {
1189         struct ptlrpc_request *req = NULL;
1190         struct obd_import *imp;
1191         int free, sent = 0;
1192         int rc = 0;
1193         ENTRY;
1194
1195         LASSERT(exp != NULL);
1196         LASSERT(count > 0);
1197
1198         CFS_FAIL_TIMEOUT(OBD_FAIL_LDLM_PAUSE_CANCEL, cfs_fail_val);
1199
1200         if (CFS_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_RACE))
1201                 RETURN(count);
1202
1203         free = ldlm_format_handles_avail(class_exp2cliimp(exp),
1204                                          &RQF_LDLM_CANCEL, RCL_CLIENT, 0);
1205         if (count > free)
1206                 count = free;
1207
1208         while (1) {
1209                 imp = class_exp2cliimp(exp);
1210                 if (imp == NULL || imp->imp_invalid) {
1211                         CDEBUG(D_DLMTRACE,
1212                                "skipping cancel on invalid import %p\n", imp);
1213                         RETURN(count);
1214                 }
1215
1216                 req = ptlrpc_request_alloc(imp, &RQF_LDLM_CANCEL);
1217                 if (req == NULL)
1218                         GOTO(out, rc = -ENOMEM);
1219
1220                 req_capsule_filled_sizes(&req->rq_pill, RCL_CLIENT);
1221                 req_capsule_set_size(&req->rq_pill, &RMF_DLM_REQ, RCL_CLIENT,
1222                                      ldlm_request_bufsize(count, LDLM_CANCEL));
1223
1224                 rc = ptlrpc_request_pack(req, LUSTRE_DLM_VERSION, LDLM_CANCEL);
1225                 if (rc) {
1226                         ptlrpc_request_free(req);
1227                         GOTO(out, rc);
1228                 }
1229
1230                 req->rq_request_portal = LDLM_CANCEL_REQUEST_PORTAL;
1231                 req->rq_reply_portal = LDLM_CANCEL_REPLY_PORTAL;
1232                 ptlrpc_at_set_req_timeout(req);
1233
1234                 ldlm_cancel_pack(req, cancels, count);
1235
1236                 ptlrpc_request_set_replen(req);
1237                 if (flags & LCF_ASYNC) {
1238                         ptlrpcd_add_req(req);
1239                         sent = count;
1240                         GOTO(out, 0);
1241                 }
1242
1243                 rc = ptlrpc_queue_wait(req);
1244                 if (rc == LUSTRE_ESTALE) {
1245                         CDEBUG(D_DLMTRACE, "client/server (nid %s) "
1246                                "out of sync -- not fatal\n",
1247                                libcfs_nid2str(req->rq_import->
1248                                               imp_connection->c_peer.nid));
1249                         rc = 0;
1250                 } else if (rc == -ETIMEDOUT && /* check there was no reconnect*/
1251                            req->rq_import_generation == imp->imp_generation) {
1252                         ptlrpc_req_finished(req);
1253                         continue;
1254                 } else if (rc != ELDLM_OK) {
1255                         /* -ESHUTDOWN is common on umount */
1256                         CDEBUG_LIMIT(rc == -ESHUTDOWN ? D_DLMTRACE : D_ERROR,
1257                                      "Got rc %d from cancel RPC: "
1258                                      "canceling anyway\n", rc);
1259                         break;
1260                 }
1261                 sent = count;
1262                 break;
1263         }
1264
1265         ptlrpc_req_finished(req);
1266         EXIT;
1267 out:
1268         return sent ? sent : rc;
1269 }
1270
1271 static inline struct ldlm_pool *ldlm_imp2pl(struct obd_import *imp)
1272 {
1273         LASSERT(imp != NULL);
1274         return &imp->imp_obd->obd_namespace->ns_pool;
1275 }
1276
1277 /**
1278  * Update client's OBD pool related fields with new SLV and Limit from \a req.
1279  */
1280 int ldlm_cli_update_pool(struct ptlrpc_request *req)
1281 {
1282         struct obd_device *obd;
1283         __u64 new_slv;
1284         __u32 new_limit;
1285         ENTRY;
1286         if (unlikely(!req->rq_import || !req->rq_import->imp_obd ||
1287                      !imp_connect_lru_resize(req->rq_import)))
1288         {
1289                 /*
1290                  * Do nothing for corner cases.
1291                  */
1292                 RETURN(0);
1293         }
1294
1295         /* In some cases RPC may contain SLV and limit zeroed out. This
1296          * is the case when server does not support LRU resize feature.
1297          * This is also possible in some recovery cases when server-side
1298          * reqs have no reference to the OBD export and thus access to
1299          * server-side namespace is not possible. */
1300         if (lustre_msg_get_slv(req->rq_repmsg) == 0 ||
1301             lustre_msg_get_limit(req->rq_repmsg) == 0) {
1302                 DEBUG_REQ(D_HA, req, "Zero SLV or Limit found "
1303                           "(SLV: "LPU64", Limit: %u)",
1304                           lustre_msg_get_slv(req->rq_repmsg),
1305                           lustre_msg_get_limit(req->rq_repmsg));
1306                 RETURN(0);
1307         }
1308
1309         new_limit = lustre_msg_get_limit(req->rq_repmsg);
1310         new_slv = lustre_msg_get_slv(req->rq_repmsg);
1311         obd = req->rq_import->imp_obd;
1312
1313         /* Set new SLV and limit in OBD fields to make them accessible
1314          * to the pool thread. We do not access obd_namespace and pool
1315          * directly here as there is no reliable way to make sure that
1316          * they are still alive at cleanup time. Evil races are possible
1317          * which may cause Oops at that time. */
1318         write_lock(&obd->obd_pool_lock);
1319         obd->obd_pool_slv = new_slv;
1320         obd->obd_pool_limit = new_limit;
1321         write_unlock(&obd->obd_pool_lock);
1322
1323         RETURN(0);
1324 }
1325
1326 /**
1327  * Client side lock cancel.
1328  *
1329  * Lock must not have any readers or writers by this time.
1330  */
1331 int ldlm_cli_cancel(struct lustre_handle *lockh,
1332                     ldlm_cancel_flags_t cancel_flags)
1333 {
1334         struct obd_export *exp;
1335         int avail, flags, count = 1;
1336         __u64 rc = 0;
1337         struct ldlm_namespace *ns;
1338         struct ldlm_lock *lock;
1339         struct list_head cancels = LIST_HEAD_INIT(cancels);
1340         ENTRY;
1341
1342         /* concurrent cancels on the same handle can happen */
1343         lock = ldlm_handle2lock_long(lockh, LDLM_FL_CANCELING);
1344         if (lock == NULL) {
1345                 LDLM_DEBUG_NOLOCK("lock is already being destroyed");
1346                 RETURN(0);
1347         }
1348
1349         rc = ldlm_cli_cancel_local(lock);
1350         if (rc == LDLM_FL_LOCAL_ONLY || cancel_flags & LCF_LOCAL) {
1351                 LDLM_LOCK_RELEASE(lock);
1352                 RETURN(0);
1353         }
1354         /* Even if the lock is marked as LDLM_FL_BL_AST, this is a LDLM_CANCEL
1355          * RPC which goes to canceld portal, so we can cancel other LRU locks
1356          * here and send them all as one LDLM_CANCEL RPC. */
1357         LASSERT(list_empty(&lock->l_bl_ast));
1358         list_add(&lock->l_bl_ast, &cancels);
1359
1360         exp = lock->l_conn_export;
1361         if (exp_connect_cancelset(exp)) {
1362                 avail = ldlm_format_handles_avail(class_exp2cliimp(exp),
1363                                                   &RQF_LDLM_CANCEL,
1364                                                   RCL_CLIENT, 0);
1365                 LASSERT(avail > 0);
1366
1367                 ns = ldlm_lock_to_ns(lock);
1368                 flags = ns_connect_lru_resize(ns) ?
1369                         LDLM_CANCEL_LRUR : LDLM_CANCEL_AGED;
1370                 count += ldlm_cancel_lru_local(ns, &cancels, 0, avail - 1,
1371                                                LCF_BL_AST, flags);
1372         }
1373         ldlm_cli_cancel_list(&cancels, count, NULL, cancel_flags);
1374         RETURN(0);
1375 }
1376 EXPORT_SYMBOL(ldlm_cli_cancel);
1377
1378 /**
1379  * Locally cancel up to \a count locks in list \a cancels.
1380  * Return the number of cancelled locks.
1381  */
1382 int ldlm_cli_cancel_list_local(struct list_head *cancels, int count,
1383                                ldlm_cancel_flags_t flags)
1384 {
1385         struct list_head head = LIST_HEAD_INIT(head);
1386         struct ldlm_lock *lock, *next;
1387         int left = 0, bl_ast = 0;
1388         __u64 rc;
1389
1390         left = count;
1391         list_for_each_entry_safe(lock, next, cancels, l_bl_ast) {
1392                 if (left-- == 0)
1393                         break;
1394
1395                 if (flags & LCF_LOCAL) {
1396                         rc = LDLM_FL_LOCAL_ONLY;
1397                         ldlm_lock_cancel(lock);
1398                 } else {
1399                         rc = ldlm_cli_cancel_local(lock);
1400                 }
1401                 /* Until we have compound requests and can send LDLM_CANCEL
1402                  * requests batched with generic RPCs, we need to send cancels
1403                  * with the LDLM_FL_BL_AST flag in a separate RPC from
1404                  * the one being generated now. */
1405                 if (!(flags & LCF_BL_AST) && (rc == LDLM_FL_BL_AST)) {
1406                         LDLM_DEBUG(lock, "Cancel lock separately");
1407                         list_del_init(&lock->l_bl_ast);
1408                         list_add(&lock->l_bl_ast, &head);
1409                         bl_ast++;
1410                         continue;
1411                 }
1412                 if (rc == LDLM_FL_LOCAL_ONLY) {
1413                         /* CANCEL RPC should not be sent to server. */
1414                         list_del_init(&lock->l_bl_ast);
1415                         LDLM_LOCK_RELEASE(lock);
1416                         count--;
1417                 }
1418         }
1419         if (bl_ast > 0) {
1420                 count -= bl_ast;
1421                 ldlm_cli_cancel_list(&head, bl_ast, NULL, 0);
1422         }
1423
1424         RETURN(count);
1425 }
1426
1427 /**
1428  * Cancel as many locks as possible w/o sending any RPCs (e.g. to write back
1429  * dirty data, to close a file, ...) or waiting for any RPCs in-flight (e.g.
1430  * readahead requests, ...)
1431  */
1432 static ldlm_policy_res_t ldlm_cancel_no_wait_policy(struct ldlm_namespace *ns,
1433                                                     struct ldlm_lock *lock,
1434                                                     int unused, int added,
1435                                                     int count)
1436 {
1437         ldlm_policy_res_t result = LDLM_POLICY_CANCEL_LOCK;
1438
1439         /* don't check added & count since we want to process all locks
1440          * from unused list.
1441          * It's fine to not take lock to access lock->l_resource since
1442          * the lock has already been granted so it won't change. */
1443         switch (lock->l_resource->lr_type) {
1444                 case LDLM_EXTENT:
1445                 case LDLM_IBITS:
1446                         if (ns->ns_cancel != NULL && ns->ns_cancel(lock) != 0)
1447                                 break;
1448                 default:
1449                         result = LDLM_POLICY_SKIP_LOCK;
1450                         lock_res_and_lock(lock);
1451                         ldlm_set_skipped(lock);
1452                         unlock_res_and_lock(lock);
1453                         break;
1454         }
1455
1456         RETURN(result);
1457 }
1458
1459 /**
1460  * Callback function for LRU-resize policy. Decides whether to keep
1461  * \a lock in LRU for current \a LRU size \a unused, added in current
1462  * scan \a added and number of locks to be preferably canceled \a count.
1463  *
1464  * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1465  *
1466  * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1467  */
1468 static ldlm_policy_res_t ldlm_cancel_lrur_policy(struct ldlm_namespace *ns,
1469                                                  struct ldlm_lock *lock,
1470                                                  int unused, int added,
1471                                                  int count)
1472 {
1473         cfs_time_t cur = cfs_time_current();
1474         struct ldlm_pool *pl = &ns->ns_pool;
1475         __u64 slv, lvf, lv;
1476         cfs_time_t la;
1477
1478         /* Stop LRU processing when we reach past @count or have checked all
1479          * locks in LRU. */
1480         if (count && added >= count)
1481                 return LDLM_POLICY_KEEP_LOCK;
1482
1483         /* Despite of the LV, It doesn't make sense to keep the lock which
1484          * is unused for ns_max_age time. */
1485         if (cfs_time_after(cfs_time_current(),
1486                            cfs_time_add(lock->l_last_used, ns->ns_max_age)))
1487                 return LDLM_POLICY_CANCEL_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         if (slv == 0 || lv < slv)
1501                 return LDLM_POLICY_KEEP_LOCK;
1502
1503         return LDLM_POLICY_CANCEL_LOCK;
1504 }
1505
1506 static ldlm_policy_res_t
1507 ldlm_cancel_lrur_no_wait_policy(struct ldlm_namespace *ns,
1508                                 struct ldlm_lock *lock,
1509                                 int unused, int added,
1510                                 int count)
1511 {
1512         ldlm_policy_res_t result;
1513
1514         result = ldlm_cancel_lrur_policy(ns, lock, unused, added, count);
1515         if (result == LDLM_POLICY_KEEP_LOCK)
1516                 return result;
1517
1518         return ldlm_cancel_no_wait_policy(ns, lock, unused, added, count);
1519 }
1520
1521 /**
1522  * Callback function for proc used policy. Makes decision whether to keep
1523  * \a lock in LRU for current \a LRU size \a unused, added in current scan \a
1524  * added and number of locks to be preferably canceled \a count.
1525  *
1526  * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1527  *
1528  * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1529  */
1530 static ldlm_policy_res_t ldlm_cancel_passed_policy(struct ldlm_namespace *ns,
1531                                                    struct ldlm_lock *lock,
1532                                                    int unused, int added,
1533                                                    int count)
1534 {
1535         /* Stop LRU processing when we reach past @count or have checked all
1536          * locks in LRU. */
1537         return (added >= count) ?
1538                 LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK;
1539 }
1540
1541 /**
1542  * Callback function for aged policy. Makes decision whether to keep \a lock in
1543  * LRU for current LRU size \a unused, added in current scan \a added and
1544  * number of locks to be preferably canceled \a count.
1545  *
1546  * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1547  *
1548  * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1549  */
1550 static ldlm_policy_res_t ldlm_cancel_aged_policy(struct ldlm_namespace *ns,
1551                                                  struct ldlm_lock *lock,
1552                                                  int unused, int added,
1553                                                  int count)
1554 {
1555         if ((added >= count) &&
1556             cfs_time_before(cfs_time_current(),
1557                             cfs_time_add(lock->l_last_used, ns->ns_max_age)))
1558                 return LDLM_POLICY_KEEP_LOCK;
1559
1560         return LDLM_POLICY_CANCEL_LOCK;
1561 }
1562
1563 /**
1564  * Callback function for default policy. Makes decision whether to keep \a lock
1565  * in LRU for current LRU size \a unused, added in current scan \a added and
1566  * number of locks to be preferably canceled \a count.
1567  *
1568  * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1569  *
1570  * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1571  */
1572 static ldlm_policy_res_t ldlm_cancel_default_policy(struct ldlm_namespace *ns,
1573                                                     struct ldlm_lock *lock,
1574                                                     int unused, int added,
1575                                                     int count)
1576 {
1577         /* Stop LRU processing when we reach past count or have checked all
1578          * locks in LRU. */
1579         return (added >= count) ?
1580                 LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK;
1581 }
1582
1583 typedef ldlm_policy_res_t (*ldlm_cancel_lru_policy_t)(struct ldlm_namespace *,
1584                                                       struct ldlm_lock *, int,
1585                                                       int, int);
1586
1587 static ldlm_cancel_lru_policy_t
1588 ldlm_cancel_lru_policy(struct ldlm_namespace *ns, int flags)
1589 {
1590         if (flags & LDLM_CANCEL_NO_WAIT)
1591                 return ldlm_cancel_no_wait_policy;
1592
1593         if (ns_connect_lru_resize(ns)) {
1594                 if (flags & LDLM_CANCEL_SHRINK)
1595                         /* We kill passed number of old locks. */
1596                         return ldlm_cancel_passed_policy;
1597                 else if (flags & LDLM_CANCEL_LRUR)
1598                         return ldlm_cancel_lrur_policy;
1599                 else if (flags & LDLM_CANCEL_PASSED)
1600                         return ldlm_cancel_passed_policy;
1601                 else if (flags & LDLM_CANCEL_LRUR_NO_WAIT)
1602                         return ldlm_cancel_lrur_no_wait_policy;
1603         } else {
1604                 if (flags & LDLM_CANCEL_AGED)
1605                         return ldlm_cancel_aged_policy;
1606         }
1607
1608         return ldlm_cancel_default_policy;
1609 }
1610
1611 /**
1612  * - Free space in LRU for \a count new locks,
1613  *   redundant unused locks are canceled locally;
1614  * - also cancel locally unused aged locks;
1615  * - do not cancel more than \a max locks;
1616  * - GET the found locks and add them into the \a cancels list.
1617  *
1618  * A client lock can be added to the l_bl_ast list only when it is
1619  * marked LDLM_FL_CANCELING. Otherwise, somebody is already doing
1620  * CANCEL.  There are the following use cases:
1621  * ldlm_cancel_resource_local(), ldlm_cancel_lru_local() and
1622  * ldlm_cli_cancel(), which check and set this flag properly. As any
1623  * attempt to cancel a lock rely on this flag, l_bl_ast list is accessed
1624  * later without any special locking.
1625  *
1626  * Calling policies for enabled LRU resize:
1627  * ----------------------------------------
1628  * flags & LDLM_CANCEL_LRUR - use LRU resize policy (SLV from server) to
1629  *                            cancel not more than \a count locks;
1630  *
1631  * flags & LDLM_CANCEL_PASSED - cancel \a count number of old locks (located at
1632  *                              the beginning of LRU list);
1633  *
1634  * flags & LDLM_CANCEL_SHRINK - cancel not more than \a count locks according to
1635  *                              memory pressre policy function;
1636  *
1637  * flags & LDLM_CANCEL_AGED - cancel \a count locks according to "aged policy".
1638  *
1639  * flags & LDLM_CANCEL_NO_WAIT - cancel as many unused locks as possible
1640  *                               (typically before replaying locks) w/o
1641  *                               sending any RPCs or waiting for any
1642  *                               outstanding RPC to complete.
1643  */
1644 static int ldlm_prepare_lru_list(struct ldlm_namespace *ns,
1645                                  struct list_head *cancels, int count, int max,
1646                                  int flags)
1647 {
1648         ldlm_cancel_lru_policy_t pf;
1649         struct ldlm_lock *lock, *next;
1650         int added = 0, unused, remained;
1651         int no_wait = flags & (LDLM_CANCEL_NO_WAIT | LDLM_CANCEL_LRUR_NO_WAIT);
1652         ENTRY;
1653
1654         spin_lock(&ns->ns_lock);
1655         unused = ns->ns_nr_unused;
1656         remained = unused;
1657
1658         if (!ns_connect_lru_resize(ns))
1659                 count += unused - ns->ns_max_unused;
1660
1661         pf = ldlm_cancel_lru_policy(ns, flags);
1662         LASSERT(pf != NULL);
1663
1664         while (!list_empty(&ns->ns_unused_list)) {
1665                 ldlm_policy_res_t result;
1666                 cfs_time_t last_use = 0;
1667
1668                 /* all unused locks */
1669                 if (remained-- <= 0)
1670                         break;
1671
1672                 /* For any flags, stop scanning if @max is reached. */
1673                 if (max && added >= max)
1674                         break;
1675
1676                 list_for_each_entry_safe(lock, next, &ns->ns_unused_list,
1677                                              l_lru) {
1678                         /* No locks which got blocking requests. */
1679                         LASSERT(!ldlm_is_bl_ast(lock));
1680
1681                         if (no_wait && ldlm_is_skipped(lock))
1682                                 /* already processed */
1683                                 continue;
1684
1685                         last_use = lock->l_last_used;
1686                         if (last_use == cfs_time_current())
1687                                 continue;
1688
1689                         /* Somebody is already doing CANCEL. No need for this
1690                          * lock in LRU, do not traverse it again. */
1691                         if (!ldlm_is_canceling(lock))
1692                                 break;
1693
1694                         ldlm_lock_remove_from_lru_nolock(lock);
1695                 }
1696                 if (&lock->l_lru == &ns->ns_unused_list)
1697                         break;
1698
1699                 LDLM_LOCK_GET(lock);
1700                 spin_unlock(&ns->ns_lock);
1701                 lu_ref_add(&lock->l_reference, __FUNCTION__, current);
1702
1703                 /* Pass the lock through the policy filter and see if it
1704                  * should stay in LRU.
1705                  *
1706                  * Even for shrinker policy we stop scanning if
1707                  * we find a lock that should stay in the cache.
1708                  * We should take into account lock age anyway
1709                  * as a new lock is a valuable resource even if
1710                  * it has a low weight.
1711                  *
1712                  * That is, for shrinker policy we drop only
1713                  * old locks, but additionally choose them by
1714                  * their weight. Big extent locks will stay in
1715                  * the cache. */
1716                 result = pf(ns, lock, unused, added, count);
1717                 if (result == LDLM_POLICY_KEEP_LOCK) {
1718                         lu_ref_del(&lock->l_reference,
1719                                    __FUNCTION__, current);
1720                         LDLM_LOCK_RELEASE(lock);
1721                         spin_lock(&ns->ns_lock);
1722                         break;
1723                 }
1724                 if (result == LDLM_POLICY_SKIP_LOCK) {
1725                         lu_ref_del(&lock->l_reference,
1726                                    __func__, current);
1727                         LDLM_LOCK_RELEASE(lock);
1728                         spin_lock(&ns->ns_lock);
1729                         continue;
1730                 }
1731
1732                 lock_res_and_lock(lock);
1733                 /* Check flags again under the lock. */
1734                 if (ldlm_is_canceling(lock) ||
1735                     ldlm_lock_remove_from_lru_check(lock, last_use) == 0) {
1736                         /* Another thread is removing lock from LRU, or
1737                          * somebody is already doing CANCEL, or there
1738                          * is a blocking request which will send cancel
1739                          * by itself, or the lock is no longer unused or
1740                          * the lock has been used since the pf() call and
1741                          * pages could be put under it. */
1742                         unlock_res_and_lock(lock);
1743                         lu_ref_del(&lock->l_reference, __FUNCTION__, current);
1744                         LDLM_LOCK_RELEASE(lock);
1745                         spin_lock(&ns->ns_lock);
1746                         continue;
1747                 }
1748                 LASSERT(!lock->l_readers && !lock->l_writers);
1749
1750                 /* If we have chosen to cancel this lock voluntarily, we
1751                  * better send cancel notification to server, so that it
1752                  * frees appropriate state. This might lead to a race
1753                  * where while we are doing cancel here, server is also
1754                  * silently cancelling this lock. */
1755                 ldlm_clear_cancel_on_block(lock);
1756
1757                 /* Setting the CBPENDING flag is a little misleading,
1758                  * but prevents an important race; namely, once
1759                  * CBPENDING is set, the lock can accumulate no more
1760                  * readers/writers. Since readers and writers are
1761                  * already zero here, ldlm_lock_decref() won't see
1762                  * this flag and call l_blocking_ast */
1763                 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_CANCELING;
1764
1765                 /* We can't re-add to l_lru as it confuses the
1766                  * refcounting in ldlm_lock_remove_from_lru() if an AST
1767                  * arrives after we drop lr_lock below. We use l_bl_ast
1768                  * and can't use l_pending_chain as it is used both on
1769                  * server and client nevertheless bug 5666 says it is
1770                  * used only on server */
1771                 LASSERT(list_empty(&lock->l_bl_ast));
1772                 list_add(&lock->l_bl_ast, cancels);
1773                 unlock_res_and_lock(lock);
1774                 lu_ref_del(&lock->l_reference, __FUNCTION__, current);
1775                 spin_lock(&ns->ns_lock);
1776                 added++;
1777                 unused--;
1778         }
1779         spin_unlock(&ns->ns_lock);
1780         RETURN(added);
1781 }
1782
1783 int ldlm_cancel_lru_local(struct ldlm_namespace *ns, struct list_head *cancels,
1784                           int count, int max, ldlm_cancel_flags_t cancel_flags,
1785                           int flags)
1786 {
1787         int added;
1788         added = ldlm_prepare_lru_list(ns, cancels, count, max, flags);
1789         if (added <= 0)
1790                 return added;
1791         return ldlm_cli_cancel_list_local(cancels, added, cancel_flags);
1792 }
1793
1794 /**
1795  * Cancel at least \a nr locks from given namespace LRU.
1796  *
1797  * When called with LCF_ASYNC the blocking callback will be handled
1798  * in a thread and this function will return after the thread has been
1799  * asked to call the callback.  When called with LCF_ASYNC the blocking
1800  * callback will be performed in this function.
1801  */
1802 int ldlm_cancel_lru(struct ldlm_namespace *ns, int nr,
1803                     ldlm_cancel_flags_t cancel_flags,
1804                     int flags)
1805 {
1806         struct list_head cancels = LIST_HEAD_INIT(cancels);
1807         int count, rc;
1808         ENTRY;
1809
1810         /* Just prepare the list of locks, do not actually cancel them yet.
1811          * Locks are cancelled later in a separate thread. */
1812         count = ldlm_prepare_lru_list(ns, &cancels, nr, 0, flags);
1813         rc = ldlm_bl_to_thread_list(ns, NULL, &cancels, count, cancel_flags);
1814         if (rc == 0)
1815                 RETURN(count);
1816
1817         RETURN(0);
1818 }
1819
1820 /**
1821  * Find and cancel locally unused locks found on resource, matched to the
1822  * given policy, mode. GET the found locks and add them into the \a cancels
1823  * list.
1824  */
1825 int ldlm_cancel_resource_local(struct ldlm_resource *res,
1826                                struct list_head *cancels,
1827                                ldlm_policy_data_t *policy,
1828                                ldlm_mode_t mode, __u64 lock_flags,
1829                                ldlm_cancel_flags_t cancel_flags, void *opaque)
1830 {
1831         struct ldlm_lock *lock;
1832         int count = 0;
1833         ENTRY;
1834
1835         lock_res(res);
1836         list_for_each_entry(lock, &res->lr_granted, l_res_link) {
1837                 if (opaque != NULL && lock->l_ast_data != opaque) {
1838                         LDLM_ERROR(lock, "data %p doesn't match opaque %p",
1839                                    lock->l_ast_data, opaque);
1840                         //LBUG();
1841                         continue;
1842                 }
1843
1844                 if (lock->l_readers || lock->l_writers)
1845                         continue;
1846
1847                 /* If somebody is already doing CANCEL, or blocking AST came,
1848                  * skip this lock. */
1849                 if (ldlm_is_bl_ast(lock) || ldlm_is_canceling(lock))
1850                         continue;
1851
1852                 if (lockmode_compat(lock->l_granted_mode, mode))
1853                         continue;
1854
1855                 /* If policy is given and this is IBITS lock, add to list only
1856                  * those locks that match by policy. */
1857                 if (policy && (lock->l_resource->lr_type == LDLM_IBITS) &&
1858                     !(lock->l_policy_data.l_inodebits.bits &
1859                       policy->l_inodebits.bits))
1860                         continue;
1861
1862                 /* See CBPENDING comment in ldlm_cancel_lru */
1863                 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_CANCELING |
1864                                  lock_flags;
1865
1866                 LASSERT(list_empty(&lock->l_bl_ast));
1867                 list_add(&lock->l_bl_ast, cancels);
1868                 LDLM_LOCK_GET(lock);
1869                 count++;
1870         }
1871         unlock_res(res);
1872
1873         RETURN(ldlm_cli_cancel_list_local(cancels, count, cancel_flags));
1874 }
1875 EXPORT_SYMBOL(ldlm_cancel_resource_local);
1876
1877 /**
1878  * Cancel client-side locks from a list and send/prepare cancel RPCs to the
1879  * server.
1880  * If \a req is NULL, send CANCEL request to server with handles of locks
1881  * in the \a cancels. If EARLY_CANCEL is not supported, send CANCEL requests
1882  * separately per lock.
1883  * If \a req is not NULL, put handles of locks in \a cancels into the request
1884  * buffer at the offset \a off.
1885  * Destroy \a cancels at the end.
1886  */
1887 int ldlm_cli_cancel_list(struct list_head *cancels, int count,
1888                          struct ptlrpc_request *req, ldlm_cancel_flags_t flags)
1889 {
1890         struct ldlm_lock *lock;
1891         int res = 0;
1892         ENTRY;
1893
1894         if (list_empty(cancels) || count == 0)
1895                 RETURN(0);
1896
1897         /* XXX: requests (both batched and not) could be sent in parallel.
1898          * Usually it is enough to have just 1 RPC, but it is possible that
1899          * there are too many locks to be cancelled in LRU or on a resource.
1900          * It would also speed up the case when the server does not support
1901          * the feature. */
1902         while (count > 0) {
1903                 LASSERT(!list_empty(cancels));
1904                 lock = list_entry(cancels->next, struct ldlm_lock,
1905                                       l_bl_ast);
1906                 LASSERT(lock->l_conn_export);
1907
1908                 if (exp_connect_cancelset(lock->l_conn_export)) {
1909                         res = count;
1910                         if (req)
1911                                 ldlm_cancel_pack(req, cancels, count);
1912                         else
1913                                 res = ldlm_cli_cancel_req(lock->l_conn_export,
1914                                                           cancels, count,
1915                                                           flags);
1916                 } else {
1917                         res = ldlm_cli_cancel_req(lock->l_conn_export,
1918                                                   cancels, 1, flags);
1919                 }
1920
1921                 if (res < 0) {
1922                         CDEBUG_LIMIT(res == -ESHUTDOWN ? D_DLMTRACE : D_ERROR,
1923                                      "ldlm_cli_cancel_list: %d\n", res);
1924                         res = count;
1925                 }
1926
1927                 count -= res;
1928                 ldlm_lock_list_put(cancels, l_bl_ast, res);
1929         }
1930         LASSERT(count == 0);
1931         RETURN(0);
1932 }
1933 EXPORT_SYMBOL(ldlm_cli_cancel_list);
1934
1935 /**
1936  * Cancel all locks on a resource that have 0 readers/writers.
1937  *
1938  * If flags & LDLM_FL_LOCAL_ONLY, throw the locks away without trying
1939  * to notify the server. */
1940 int ldlm_cli_cancel_unused_resource(struct ldlm_namespace *ns,
1941                                     const struct ldlm_res_id *res_id,
1942                                     ldlm_policy_data_t *policy,
1943                                     ldlm_mode_t mode,
1944                                     ldlm_cancel_flags_t flags,
1945                                     void *opaque)
1946 {
1947         struct ldlm_resource *res;
1948         struct list_head cancels = LIST_HEAD_INIT(cancels);
1949         int count;
1950         int rc;
1951         ENTRY;
1952
1953         res = ldlm_resource_get(ns, NULL, res_id, 0, 0);
1954         if (IS_ERR(res)) {
1955                 /* This is not a problem. */
1956                 CDEBUG(D_INFO, "No resource "LPU64"\n", res_id->name[0]);
1957                 RETURN(0);
1958         }
1959
1960         LDLM_RESOURCE_ADDREF(res);
1961         count = ldlm_cancel_resource_local(res, &cancels, policy, mode,
1962                                            0, flags | LCF_BL_AST, opaque);
1963         rc = ldlm_cli_cancel_list(&cancels, count, NULL, flags);
1964         if (rc != ELDLM_OK)
1965                 CERROR("canceling unused lock "DLDLMRES": rc = %d\n",
1966                        PLDLMRES(res), rc);
1967
1968         LDLM_RESOURCE_DELREF(res);
1969         ldlm_resource_putref(res);
1970         RETURN(0);
1971 }
1972 EXPORT_SYMBOL(ldlm_cli_cancel_unused_resource);
1973
1974 struct ldlm_cli_cancel_arg {
1975         int     lc_flags;
1976         void   *lc_opaque;
1977 };
1978
1979 static int
1980 ldlm_cli_hash_cancel_unused(struct cfs_hash *hs, struct cfs_hash_bd *bd,
1981                             struct hlist_node *hnode, void *arg)
1982 {
1983         struct ldlm_resource           *res = cfs_hash_object(hs, hnode);
1984         struct ldlm_cli_cancel_arg     *lc = arg;
1985
1986         ldlm_cli_cancel_unused_resource(ldlm_res_to_ns(res), &res->lr_name,
1987                                         NULL, LCK_MINMODE, lc->lc_flags,
1988                                         lc->lc_opaque);
1989         /* must return 0 for hash iteration */
1990         return 0;
1991 }
1992
1993 /**
1994  * Cancel all locks on a namespace (or a specific resource, if given)
1995  * that have 0 readers/writers.
1996  *
1997  * If flags & LCF_LOCAL, throw the locks away without trying
1998  * to notify the server. */
1999 int ldlm_cli_cancel_unused(struct ldlm_namespace *ns,
2000                            const struct ldlm_res_id *res_id,
2001                            ldlm_cancel_flags_t flags, void *opaque)
2002 {
2003         struct ldlm_cli_cancel_arg arg = {
2004                 .lc_flags       = flags,
2005                 .lc_opaque      = opaque,
2006         };
2007
2008         ENTRY;
2009
2010         if (ns == NULL)
2011                 RETURN(ELDLM_OK);
2012
2013         if (res_id != NULL) {
2014                 RETURN(ldlm_cli_cancel_unused_resource(ns, res_id, NULL,
2015                                                        LCK_MINMODE, flags,
2016                                                        opaque));
2017         } else {
2018                 cfs_hash_for_each_nolock(ns->ns_rs_hash,
2019                                          ldlm_cli_hash_cancel_unused, &arg);
2020                 RETURN(ELDLM_OK);
2021         }
2022 }
2023
2024 /* Lock iterators. */
2025
2026 int ldlm_resource_foreach(struct ldlm_resource *res, ldlm_iterator_t iter,
2027                           void *closure)
2028 {
2029         struct list_head *tmp, *next;
2030         struct ldlm_lock *lock;
2031         int rc = LDLM_ITER_CONTINUE;
2032
2033         ENTRY;
2034
2035         if (!res)
2036                 RETURN(LDLM_ITER_CONTINUE);
2037
2038         lock_res(res);
2039         list_for_each_safe(tmp, next, &res->lr_granted) {
2040                 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
2041
2042                 if (iter(lock, closure) == LDLM_ITER_STOP)
2043                         GOTO(out, rc = LDLM_ITER_STOP);
2044         }
2045
2046         list_for_each_safe(tmp, next, &res->lr_converting) {
2047                 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
2048
2049                 if (iter(lock, closure) == LDLM_ITER_STOP)
2050                         GOTO(out, rc = LDLM_ITER_STOP);
2051         }
2052
2053         list_for_each_safe(tmp, next, &res->lr_waiting) {
2054                 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
2055
2056                 if (iter(lock, closure) == LDLM_ITER_STOP)
2057                         GOTO(out, rc = LDLM_ITER_STOP);
2058         }
2059  out:
2060         unlock_res(res);
2061         RETURN(rc);
2062 }
2063
2064 struct iter_helper_data {
2065         ldlm_iterator_t iter;
2066         void *closure;
2067 };
2068
2069 static int ldlm_iter_helper(struct ldlm_lock *lock, void *closure)
2070 {
2071         struct iter_helper_data *helper = closure;
2072         return helper->iter(lock, helper->closure);
2073 }
2074
2075 static int ldlm_res_iter_helper(struct cfs_hash *hs, struct cfs_hash_bd *bd,
2076                                 struct hlist_node *hnode, void *arg)
2077
2078 {
2079         struct ldlm_resource *res = cfs_hash_object(hs, hnode);
2080
2081         return ldlm_resource_foreach(res, ldlm_iter_helper, arg) ==
2082                LDLM_ITER_STOP;
2083 }
2084
2085 void ldlm_namespace_foreach(struct ldlm_namespace *ns,
2086                             ldlm_iterator_t iter, void *closure)
2087
2088 {
2089         struct iter_helper_data helper = { .iter = iter, .closure = closure };
2090
2091         cfs_hash_for_each_nolock(ns->ns_rs_hash,
2092                                  ldlm_res_iter_helper, &helper);
2093
2094 }
2095
2096 /* non-blocking function to manipulate a lock whose cb_data is being put away.
2097  * return  0:  find no resource
2098  *       > 0:  must be LDLM_ITER_STOP/LDLM_ITER_CONTINUE.
2099  *       < 0:  errors
2100  */
2101 int ldlm_resource_iterate(struct ldlm_namespace *ns,
2102                           const struct ldlm_res_id *res_id,
2103                           ldlm_iterator_t iter, void *data)
2104 {
2105         struct ldlm_resource *res;
2106         int rc;
2107         ENTRY;
2108
2109         LASSERTF(ns != NULL, "must pass in namespace\n");
2110
2111         res = ldlm_resource_get(ns, NULL, res_id, 0, 0);
2112         if (IS_ERR(res))
2113                 RETURN(0);
2114
2115         LDLM_RESOURCE_ADDREF(res);
2116         rc = ldlm_resource_foreach(res, iter, data);
2117         LDLM_RESOURCE_DELREF(res);
2118         ldlm_resource_putref(res);
2119         RETURN(rc);
2120 }
2121 EXPORT_SYMBOL(ldlm_resource_iterate);
2122
2123 /* Lock replay */
2124
2125 static int ldlm_chain_lock_for_replay(struct ldlm_lock *lock, void *closure)
2126 {
2127         struct list_head *list = closure;
2128
2129         /* we use l_pending_chain here, because it's unused on clients. */
2130         LASSERTF(list_empty(&lock->l_pending_chain),
2131                  "lock %p next %p prev %p\n",
2132                  lock, &lock->l_pending_chain.next,&lock->l_pending_chain.prev);
2133         /* bug 9573: don't replay locks left after eviction, or
2134          * bug 17614: locks being actively cancelled. Get a reference
2135          * on a lock so that it does not disapear under us (e.g. due to cancel)
2136          */
2137         if (!(lock->l_flags & (LDLM_FL_FAILED|LDLM_FL_CANCELING))) {
2138                 list_add(&lock->l_pending_chain, list);
2139                 LDLM_LOCK_GET(lock);
2140         }
2141
2142         return LDLM_ITER_CONTINUE;
2143 }
2144
2145 static int replay_lock_interpret(const struct lu_env *env,
2146                                  struct ptlrpc_request *req,
2147                                  struct ldlm_async_args *aa, int rc)
2148 {
2149         struct ldlm_lock     *lock;
2150         struct ldlm_reply    *reply;
2151         struct obd_export    *exp;
2152
2153         ENTRY;
2154         atomic_dec(&req->rq_import->imp_replay_inflight);
2155         if (rc != ELDLM_OK)
2156                 GOTO(out, rc);
2157
2158         reply = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
2159         if (reply == NULL)
2160                 GOTO(out, rc = -EPROTO);
2161
2162         lock = ldlm_handle2lock(&aa->lock_handle);
2163         if (!lock) {
2164                 CERROR("received replay ack for unknown local cookie "LPX64
2165                        " remote cookie "LPX64 " from server %s id %s\n",
2166                        aa->lock_handle.cookie, reply->lock_handle.cookie,
2167                        req->rq_export->exp_client_uuid.uuid,
2168                        libcfs_id2str(req->rq_peer));
2169                 GOTO(out, rc = -ESTALE);
2170         }
2171
2172         /* Key change rehash lock in per-export hash with new key */
2173         exp = req->rq_export;
2174         if (exp && exp->exp_lock_hash) {
2175                 /* In the function below, .hs_keycmp resolves to
2176                  * ldlm_export_lock_keycmp() */
2177                 /* coverity[overrun-buffer-val] */
2178                 cfs_hash_rehash_key(exp->exp_lock_hash,
2179                                     &lock->l_remote_handle,
2180                                     &reply->lock_handle,
2181                                     &lock->l_exp_hash);
2182         } else {
2183                 lock->l_remote_handle = reply->lock_handle;
2184         }
2185
2186         LDLM_DEBUG(lock, "replayed lock:");
2187         ptlrpc_import_recovery_state_machine(req->rq_import);
2188         LDLM_LOCK_PUT(lock);
2189 out:
2190         if (rc != ELDLM_OK)
2191                 ptlrpc_connect_import(req->rq_import);
2192
2193         RETURN(rc);
2194 }
2195
2196 static int replay_one_lock(struct obd_import *imp, struct ldlm_lock *lock)
2197 {
2198         struct ptlrpc_request *req;
2199         struct ldlm_async_args *aa;
2200         struct ldlm_request   *body;
2201         int flags;
2202         ENTRY;
2203
2204
2205         /* Bug 11974: Do not replay a lock which is actively being canceled */
2206         if (ldlm_is_canceling(lock)) {
2207                 LDLM_DEBUG(lock, "Not replaying canceled lock:");
2208                 RETURN(0);
2209         }
2210
2211         /* If this is reply-less callback lock, we cannot replay it, since
2212          * server might have long dropped it, but notification of that event was
2213          * lost by network. (and server granted conflicting lock already) */
2214         if (ldlm_is_cancel_on_block(lock)) {
2215                 LDLM_DEBUG(lock, "Not replaying reply-less lock:");
2216                 ldlm_lock_cancel(lock);
2217                 RETURN(0);
2218         }
2219
2220         /*
2221          * If granted mode matches the requested mode, this lock is granted.
2222          *
2223          * If they differ, but we have a granted mode, then we were granted
2224          * one mode and now want another: ergo, converting.
2225          *
2226          * If we haven't been granted anything and are on a resource list,
2227          * then we're blocked/waiting.
2228          *
2229          * If we haven't been granted anything and we're NOT on a resource list,
2230          * then we haven't got a reply yet and don't have a known disposition.
2231          * This happens whenever a lock enqueue is the request that triggers
2232          * recovery.
2233          */
2234         if (lock->l_granted_mode == lock->l_req_mode)
2235                 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_GRANTED;
2236         else if (lock->l_granted_mode)
2237                 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_CONV;
2238         else if (!list_empty(&lock->l_res_link))
2239                 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_WAIT;
2240         else
2241                 flags = LDLM_FL_REPLAY;
2242
2243         req = ptlrpc_request_alloc_pack(imp, &RQF_LDLM_ENQUEUE,
2244                                         LUSTRE_DLM_VERSION, LDLM_ENQUEUE);
2245         if (req == NULL)
2246                 RETURN(-ENOMEM);
2247
2248         /* We're part of recovery, so don't wait for it. */
2249         req->rq_send_state = LUSTRE_IMP_REPLAY_LOCKS;
2250
2251         body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2252         ldlm_lock2desc(lock, &body->lock_desc);
2253         body->lock_flags = ldlm_flags_to_wire(flags);
2254
2255         ldlm_lock2handle(lock, &body->lock_handle[0]);
2256         if (lock->l_lvb_len > 0)
2257                 req_capsule_extend(&req->rq_pill, &RQF_LDLM_ENQUEUE_LVB);
2258         req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
2259                              lock->l_lvb_len);
2260         ptlrpc_request_set_replen(req);
2261         /* notify the server we've replayed all requests.
2262          * also, we mark the request to be put on a dedicated
2263          * queue to be processed after all request replayes.
2264          * bug 6063 */
2265         lustre_msg_set_flags(req->rq_reqmsg, MSG_REQ_REPLAY_DONE);
2266
2267         LDLM_DEBUG(lock, "replaying lock:");
2268
2269         atomic_inc(&req->rq_import->imp_replay_inflight);
2270         CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
2271         aa = ptlrpc_req_async_args(req);
2272         aa->lock_handle = body->lock_handle[0];
2273         req->rq_interpret_reply = (ptlrpc_interpterer_t)replay_lock_interpret;
2274         ptlrpcd_add_req(req);
2275
2276         RETURN(0);
2277 }
2278
2279 /**
2280  * Cancel as many unused locks as possible before replay. since we are
2281  * in recovery, we can't wait for any outstanding RPCs to send any RPC
2282  * to the server.
2283  *
2284  * Called only in recovery before replaying locks. there is no need to
2285  * replay locks that are unused. since the clients may hold thousands of
2286  * cached unused locks, dropping the unused locks can greatly reduce the
2287  * load on the servers at recovery time.
2288  */
2289 static void ldlm_cancel_unused_locks_for_replay(struct ldlm_namespace *ns)
2290 {
2291         int canceled;
2292         struct list_head cancels = LIST_HEAD_INIT(cancels);
2293
2294         CDEBUG(D_DLMTRACE, "Dropping as many unused locks as possible before"
2295                            "replay for namespace %s (%d)\n",
2296                            ldlm_ns_name(ns), ns->ns_nr_unused);
2297
2298         /* We don't need to care whether or not LRU resize is enabled
2299          * because the LDLM_CANCEL_NO_WAIT policy doesn't use the
2300          * count parameter */
2301         canceled = ldlm_cancel_lru_local(ns, &cancels, ns->ns_nr_unused, 0,
2302                                          LCF_LOCAL, LDLM_CANCEL_NO_WAIT);
2303
2304         CDEBUG(D_DLMTRACE, "Canceled %d unused locks from namespace %s\n",
2305                            canceled, ldlm_ns_name(ns));
2306 }
2307
2308 int ldlm_replay_locks(struct obd_import *imp)
2309 {
2310         struct ldlm_namespace *ns = imp->imp_obd->obd_namespace;
2311         struct list_head list = LIST_HEAD_INIT(list);
2312         struct ldlm_lock *lock, *next;
2313         int rc = 0;
2314
2315         ENTRY;
2316
2317         LASSERT(atomic_read(&imp->imp_replay_inflight) == 0);
2318
2319         /* don't replay locks if import failed recovery */
2320         if (imp->imp_vbr_failed)
2321                 RETURN(0);
2322
2323         /* ensure this doesn't fall to 0 before all have been queued */
2324         atomic_inc(&imp->imp_replay_inflight);
2325
2326         if (ldlm_cancel_unused_locks_before_replay)
2327                 ldlm_cancel_unused_locks_for_replay(ns);
2328
2329         ldlm_namespace_foreach(ns, ldlm_chain_lock_for_replay, &list);
2330
2331         list_for_each_entry_safe(lock, next, &list, l_pending_chain) {
2332                 list_del_init(&lock->l_pending_chain);
2333                 if (rc) {
2334                         LDLM_LOCK_RELEASE(lock);
2335                         continue; /* or try to do the rest? */
2336                 }
2337                 rc = replay_one_lock(imp, lock);
2338                 LDLM_LOCK_RELEASE(lock);
2339         }
2340
2341         atomic_dec(&imp->imp_replay_inflight);
2342
2343         RETURN(rc);
2344 }