4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
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.
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).
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
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
27 * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
30 * Copyright (c) 2010, 2012, Intel Corporation.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
37 * This file contains Asynchronous System Trap (AST) handlers and related
38 * LDLM request-processing routines.
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:
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
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.
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
62 #define DEBUG_SUBSYSTEM S_LDLM
64 #include <lustre_dlm.h>
65 #include <obd_class.h>
68 #include "ldlm_internal.h"
70 int ldlm_enqueue_min
= OBD_TIMEOUT_DEFAULT
;
71 CFS_MODULE_PARM(ldlm_enqueue_min
, "i", int, 0644,
72 "lock enqueue timeout minimum");
74 /* in client side, whether the cached locks will be canceled before replay */
75 unsigned int ldlm_cancel_unused_locks_before_replay
= 1;
77 static void interrupted_completion_wait(void *data
)
81 struct lock_wait_data
{
82 struct ldlm_lock
*lwd_lock
;
86 struct ldlm_async_args
{
87 struct lustre_handle lock_handle
;
90 int ldlm_expired_completion_wait(void *data
)
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
;
97 if (lock
->l_conn_export
== NULL
) {
98 static cfs_time_t next_dump
= 0, last_dump
= 0;
100 if (ptlrpc_check_suspend())
103 LCONSOLE_WARN("lock timed out (enqueued at "CFS_TIME_T
", "
104 CFS_DURATION_T
"s ago)\n",
105 lock
->l_last_activity
,
106 cfs_time_sub(cfs_time_current_sec(),
107 lock
->l_last_activity
));
108 LDLM_DEBUG(lock
, "lock timed out (enqueued at "CFS_TIME_T
", "
109 CFS_DURATION_T
"s ago); not entering recovery in "
110 "server code, just going back to sleep",
111 lock
->l_last_activity
,
112 cfs_time_sub(cfs_time_current_sec(),
113 lock
->l_last_activity
));
114 if (cfs_time_after(cfs_time_current(), next_dump
)) {
115 last_dump
= next_dump
;
116 next_dump
= cfs_time_shift(300);
117 ldlm_namespace_dump(D_DLMTRACE
,
118 ldlm_lock_to_ns(lock
));
120 libcfs_debug_dumplog();
125 obd
= lock
->l_conn_export
->exp_obd
;
126 imp
= obd
->u
.cli
.cl_import
;
127 ptlrpc_fail_import(imp
, lwd
->lwd_conn_cnt
);
128 LDLM_ERROR(lock
, "lock timed out (enqueued at "CFS_TIME_T
", "
129 CFS_DURATION_T
"s ago), entering recovery for %s@%s",
130 lock
->l_last_activity
,
131 cfs_time_sub(cfs_time_current_sec(), lock
->l_last_activity
),
132 obd2cli_tgt(obd
), imp
->imp_connection
->c_remote_uuid
.uuid
);
136 EXPORT_SYMBOL(ldlm_expired_completion_wait
);
138 /* We use the same basis for both server side and client side functions
139 from a single node. */
140 int ldlm_get_enq_timeout(struct ldlm_lock
*lock
)
142 int timeout
= at_get(ldlm_lock_to_ns_at(lock
));
144 return obd_timeout
/ 2;
145 /* Since these are non-updating timeouts, we should be conservative.
146 It would be nice to have some kind of "early reply" mechanism for
147 lock callbacks too... */
148 timeout
= min_t(int, at_max
, timeout
+ (timeout
>> 1)); /* 150% */
149 return max(timeout
, ldlm_enqueue_min
);
151 EXPORT_SYMBOL(ldlm_get_enq_timeout
);
154 * Helper function for ldlm_completion_ast(), updating timings when lock is
157 static int ldlm_completion_tail(struct ldlm_lock
*lock
)
162 if (lock
->l_flags
& (LDLM_FL_DESTROYED
| LDLM_FL_FAILED
)) {
163 LDLM_DEBUG(lock
, "client-side enqueue: destroyed");
166 delay
= cfs_time_sub(cfs_time_current_sec(),
167 lock
->l_last_activity
);
168 LDLM_DEBUG(lock
, "client-side enqueue: granted after "
169 CFS_DURATION_T
"s", delay
);
171 /* Update our time estimate */
172 at_measured(ldlm_lock_to_ns_at(lock
),
180 * Implementation of ->l_completion_ast() for a client, that doesn't wait
181 * until lock is granted. Suitable for locks enqueued through ptlrpcd, of
182 * other threads that cannot block for long.
184 int ldlm_completion_ast_async(struct ldlm_lock
*lock
, __u64 flags
, void *data
)
186 if (flags
== LDLM_FL_WAIT_NOREPROC
) {
187 LDLM_DEBUG(lock
, "client-side enqueue waiting on pending lock");
191 if (!(flags
& (LDLM_FL_BLOCK_WAIT
| LDLM_FL_BLOCK_GRANTED
|
192 LDLM_FL_BLOCK_CONV
))) {
193 wake_up(&lock
->l_waitq
);
194 return ldlm_completion_tail(lock
);
197 LDLM_DEBUG(lock
, "client-side enqueue returned a blocked lock, "
199 ldlm_reprocess_all(lock
->l_resource
);
202 EXPORT_SYMBOL(ldlm_completion_ast_async
);
205 * Generic LDLM "completion" AST. This is called in several cases:
207 * - when a reply to an ENQUEUE RPC is received from the server
208 * (ldlm_cli_enqueue_fini()). Lock might be granted or not granted at
209 * this point (determined by flags);
211 * - when LDLM_CP_CALLBACK RPC comes to client to notify it that lock has
214 * - when ldlm_lock_match(LDLM_FL_LVB_READY) is about to wait until lock
217 * - to force all locks when resource is destroyed (cleanup_resource());
219 * - during lock conversion (not used currently).
221 * If lock is not granted in the first case, this function waits until second
222 * or penultimate cases happen in some other thread.
225 int ldlm_completion_ast(struct ldlm_lock
*lock
, __u64 flags
, void *data
)
227 /* XXX ALLOCATE - 160 bytes */
228 struct lock_wait_data lwd
;
229 struct obd_device
*obd
;
230 struct obd_import
*imp
= NULL
;
231 struct l_wait_info lwi
;
235 if (flags
== LDLM_FL_WAIT_NOREPROC
) {
236 LDLM_DEBUG(lock
, "client-side enqueue waiting on pending lock");
240 if (!(flags
& (LDLM_FL_BLOCK_WAIT
| LDLM_FL_BLOCK_GRANTED
|
241 LDLM_FL_BLOCK_CONV
))) {
242 wake_up(&lock
->l_waitq
);
246 LDLM_DEBUG(lock
, "client-side enqueue returned a blocked lock, "
251 obd
= class_exp2obd(lock
->l_conn_export
);
253 /* if this is a local lock, then there is no import */
255 imp
= obd
->u
.cli
.cl_import
;
258 /* Wait a long time for enqueue - server may have to callback a
259 lock from another client. Server will evict the other client if it
260 doesn't respond reasonably, and then give us the lock. */
261 timeout
= ldlm_get_enq_timeout(lock
) * 2;
265 if (lock
->l_flags
& LDLM_FL_NO_TIMEOUT
) {
266 LDLM_DEBUG(lock
, "waiting indefinitely because of NO_TIMEOUT");
267 lwi
= LWI_INTR(interrupted_completion_wait
, &lwd
);
269 lwi
= LWI_TIMEOUT_INTR(cfs_time_seconds(timeout
),
270 ldlm_expired_completion_wait
,
271 interrupted_completion_wait
, &lwd
);
275 spin_lock(&imp
->imp_lock
);
276 lwd
.lwd_conn_cnt
= imp
->imp_conn_cnt
;
277 spin_unlock(&imp
->imp_lock
);
280 if (ns_is_client(ldlm_lock_to_ns(lock
)) &&
281 OBD_FAIL_CHECK_RESET(OBD_FAIL_LDLM_INTR_CP_AST
,
282 OBD_FAIL_LDLM_CP_BL_RACE
| OBD_FAIL_ONCE
)) {
283 lock
->l_flags
|= LDLM_FL_FAIL_LOC
;
286 /* Go to sleep until the lock is granted or cancelled. */
287 rc
= l_wait_event(lock
->l_waitq
,
288 is_granted_or_cancelled(lock
), &lwi
);
292 LDLM_DEBUG(lock
, "client-side enqueue waking up: failed (%d)",
297 return ldlm_completion_tail(lock
);
299 EXPORT_SYMBOL(ldlm_completion_ast
);
302 * A helper to build a blocking AST function
304 * Perform a common operation for blocking ASTs:
305 * defferred lock cancellation.
307 * \param lock the lock blocking or canceling AST was called on
309 * \see mdt_blocking_ast
310 * \see ldlm_blocking_ast
312 int ldlm_blocking_ast_nocheck(struct ldlm_lock
*lock
)
316 lock
->l_flags
|= LDLM_FL_CBPENDING
;
317 do_ast
= (!lock
->l_readers
&& !lock
->l_writers
);
318 unlock_res_and_lock(lock
);
321 struct lustre_handle lockh
;
324 LDLM_DEBUG(lock
, "already unused, calling ldlm_cli_cancel");
325 ldlm_lock2handle(lock
, &lockh
);
326 rc
= ldlm_cli_cancel(&lockh
, LCF_ASYNC
);
328 CERROR("ldlm_cli_cancel: %d\n", rc
);
330 LDLM_DEBUG(lock
, "Lock still has references, will be "
335 EXPORT_SYMBOL(ldlm_blocking_ast_nocheck
);
338 * Server blocking AST
340 * ->l_blocking_ast() callback for LDLM locks acquired by server-side
343 * \param lock the lock which blocks a request or cancelling lock
346 * \param flag indicates whether this cancelling or blocking callback
348 * \see ldlm_blocking_ast_nocheck
350 int ldlm_blocking_ast(struct ldlm_lock
*lock
, struct ldlm_lock_desc
*desc
,
351 void *data
, int flag
)
353 if (flag
== LDLM_CB_CANCELING
) {
354 /* Don't need to do anything here. */
358 lock_res_and_lock(lock
);
359 /* Get this: if ldlm_blocking_ast is racing with intent_policy, such
360 * that ldlm_blocking_ast is called just before intent_policy method
361 * takes the lr_lock, then by the time we get the lock, we might not
362 * be the correct blocking function anymore. So check, and return
364 if (lock
->l_blocking_ast
!= ldlm_blocking_ast
) {
365 unlock_res_and_lock(lock
);
368 return ldlm_blocking_ast_nocheck(lock
);
370 EXPORT_SYMBOL(ldlm_blocking_ast
);
373 * ->l_glimpse_ast() for DLM extent locks acquired on the server-side. See
374 * comment in filter_intent_policy() on why you may need this.
376 int ldlm_glimpse_ast(struct ldlm_lock
*lock
, void *reqp
)
379 * Returning -ELDLM_NO_LOCK_DATA actually works, but the reason for
380 * that is rather subtle: with OST-side locking, it may so happen that
381 * _all_ extent locks are held by the OST. If client wants to obtain
382 * current file size it calls ll{,u}_glimpse_size(), and (as locks are
383 * on the server), dummy glimpse callback fires and does
384 * nothing. Client still receives correct file size due to the
385 * following fragment in filter_intent_policy():
387 * rc = l->l_glimpse_ast(l, NULL); // this will update the LVB
388 * if (rc != 0 && res->lr_namespace->ns_lvbo &&
389 * res->lr_namespace->ns_lvbo->lvbo_update) {
390 * res->lr_namespace->ns_lvbo->lvbo_update(res, NULL, 0, 1);
393 * that is, after glimpse_ast() fails, filter_lvbo_update() runs, and
394 * returns correct file size to the client.
396 return -ELDLM_NO_LOCK_DATA
;
398 EXPORT_SYMBOL(ldlm_glimpse_ast
);
401 * Enqueue a local lock (typically on a server).
403 int ldlm_cli_enqueue_local(struct ldlm_namespace
*ns
,
404 const struct ldlm_res_id
*res_id
,
405 ldlm_type_t type
, ldlm_policy_data_t
*policy
,
406 ldlm_mode_t mode
, __u64
*flags
,
407 ldlm_blocking_callback blocking
,
408 ldlm_completion_callback completion
,
409 ldlm_glimpse_callback glimpse
,
410 void *data
, __u32 lvb_len
, enum lvb_type lvb_type
,
411 const __u64
*client_cookie
,
412 struct lustre_handle
*lockh
)
414 struct ldlm_lock
*lock
;
416 const struct ldlm_callback_suite cbs
= { .lcs_completion
= completion
,
417 .lcs_blocking
= blocking
,
418 .lcs_glimpse
= glimpse
,
421 LASSERT(!(*flags
& LDLM_FL_REPLAY
));
422 if (unlikely(ns_is_client(ns
))) {
423 CERROR("Trying to enqueue local lock in a shadow namespace\n");
427 lock
= ldlm_lock_create(ns
, res_id
, type
, mode
, &cbs
, data
, lvb_len
,
430 GOTO(out_nolock
, err
= -ENOMEM
);
432 ldlm_lock2handle(lock
, lockh
);
434 /* NB: we don't have any lock now (lock_res_and_lock)
435 * because it's a new lock */
436 ldlm_lock_addref_internal_nolock(lock
, mode
);
437 lock
->l_flags
|= LDLM_FL_LOCAL
;
438 if (*flags
& LDLM_FL_ATOMIC_CB
)
439 lock
->l_flags
|= LDLM_FL_ATOMIC_CB
;
442 lock
->l_policy_data
= *policy
;
443 if (client_cookie
!= NULL
)
444 lock
->l_client_cookie
= *client_cookie
;
445 if (type
== LDLM_EXTENT
)
446 lock
->l_req_extent
= policy
->l_extent
;
448 err
= ldlm_lock_enqueue(ns
, &lock
, policy
, flags
);
449 if (unlikely(err
!= ELDLM_OK
))
453 *policy
= lock
->l_policy_data
;
455 if (lock
->l_completion_ast
)
456 lock
->l_completion_ast(lock
, *flags
, NULL
);
458 LDLM_DEBUG(lock
, "client-side local enqueue handler, new lock created");
460 LDLM_LOCK_RELEASE(lock
);
464 EXPORT_SYMBOL(ldlm_cli_enqueue_local
);
466 static void failed_lock_cleanup(struct ldlm_namespace
*ns
,
467 struct ldlm_lock
*lock
, int mode
)
471 /* Set a flag to prevent us from sending a CANCEL (bug 407) */
472 lock_res_and_lock(lock
);
473 /* Check that lock is not granted or failed, we might race. */
474 if ((lock
->l_req_mode
!= lock
->l_granted_mode
) &&
475 !(lock
->l_flags
& LDLM_FL_FAILED
)) {
476 /* Make sure that this lock will not be found by raced
477 * bl_ast and -EINVAL reply is sent to server anyways.
479 lock
->l_flags
|= LDLM_FL_LOCAL_ONLY
| LDLM_FL_FAILED
|
480 LDLM_FL_ATOMIC_CB
| LDLM_FL_CBPENDING
;
483 unlock_res_and_lock(lock
);
487 "setting FL_LOCAL_ONLY | LDLM_FL_FAILED | "
488 "LDLM_FL_ATOMIC_CB | LDLM_FL_CBPENDING");
490 LDLM_DEBUG(lock
, "lock was granted or failed in race");
492 ldlm_lock_decref_internal(lock
, mode
);
494 /* XXX - HACK because we shouldn't call ldlm_lock_destroy()
495 * from llite/file.c/ll_file_flock(). */
496 /* This code makes for the fact that we do not have blocking handler on
497 * a client for flock locks. As such this is the place where we must
498 * completely kill failed locks. (interrupted and those that
499 * were waiting to be granted when server evicted us. */
500 if (lock
->l_resource
->lr_type
== LDLM_FLOCK
) {
501 lock_res_and_lock(lock
);
502 ldlm_resource_unlink_lock(lock
);
503 ldlm_lock_destroy_nolock(lock
);
504 unlock_res_and_lock(lock
);
509 * Finishing portion of client lock enqueue code.
511 * Called after receiving reply from server.
513 int ldlm_cli_enqueue_fini(struct obd_export
*exp
, struct ptlrpc_request
*req
,
514 ldlm_type_t type
, __u8 with_policy
, ldlm_mode_t mode
,
515 __u64
*flags
, void *lvb
, __u32 lvb_len
,
516 struct lustre_handle
*lockh
,int rc
)
518 struct ldlm_namespace
*ns
= exp
->exp_obd
->obd_namespace
;
519 int is_replay
= *flags
& LDLM_FL_REPLAY
;
520 struct ldlm_lock
*lock
;
521 struct ldlm_reply
*reply
;
522 int cleanup_phase
= 1;
525 lock
= ldlm_handle2lock(lockh
);
526 /* ldlm_cli_enqueue is holding a reference on this lock. */
528 LASSERT(type
== LDLM_FLOCK
);
532 LASSERTF(ergo(lvb_len
!= 0, lvb_len
== lock
->l_lvb_len
),
533 "lvb_len = %d, l_lvb_len = %d\n", lvb_len
, lock
->l_lvb_len
);
535 if (rc
!= ELDLM_OK
) {
537 LDLM_DEBUG(lock
, "client-side enqueue END (%s)",
538 rc
== ELDLM_LOCK_ABORTED
? "ABORTED" : "FAILED");
540 if (rc
!= ELDLM_LOCK_ABORTED
)
544 /* Before we return, swab the reply */
545 reply
= req_capsule_server_get(&req
->rq_pill
, &RMF_DLM_REP
);
547 GOTO(cleanup
, rc
= -EPROTO
);
550 LASSERT(lvb
!= NULL
);
552 size
= req_capsule_get_size(&req
->rq_pill
, &RMF_DLM_LVB
,
555 LDLM_ERROR(lock
, "Fail to get lvb_len, rc = %d", size
);
556 GOTO(cleanup
, rc
= size
);
557 } else if (unlikely(size
> lvb_len
)) {
558 LDLM_ERROR(lock
, "Replied LVB is larger than "
559 "expectation, expected = %d, replied = %d",
561 GOTO(cleanup
, rc
= -EINVAL
);
565 if (rc
== ELDLM_LOCK_ABORTED
) {
567 rc
= ldlm_fill_lvb(lock
, &req
->rq_pill
, RCL_SERVER
,
569 GOTO(cleanup
, rc
= (rc
!= 0 ? rc
: ELDLM_LOCK_ABORTED
));
572 /* lock enqueued on the server */
575 lock_res_and_lock(lock
);
576 /* Key change rehash lock in per-export hash with new key */
577 if (exp
->exp_lock_hash
) {
578 /* In the function below, .hs_keycmp resolves to
579 * ldlm_export_lock_keycmp() */
580 /* coverity[overrun-buffer-val] */
581 cfs_hash_rehash_key(exp
->exp_lock_hash
,
582 &lock
->l_remote_handle
,
586 lock
->l_remote_handle
= reply
->lock_handle
;
589 *flags
= ldlm_flags_from_wire(reply
->lock_flags
);
590 lock
->l_flags
|= ldlm_flags_from_wire(reply
->lock_flags
&
592 /* move NO_TIMEOUT flag to the lock to force ldlm_lock_match()
593 * to wait with no timeout as well */
594 lock
->l_flags
|= ldlm_flags_from_wire(reply
->lock_flags
&
596 unlock_res_and_lock(lock
);
598 CDEBUG(D_INFO
, "local: %p, remote cookie: "LPX64
", flags: 0x%llx\n",
599 lock
, reply
->lock_handle
.cookie
, *flags
);
601 /* If enqueue returned a blocked lock but the completion handler has
602 * already run, then it fixed up the resource and we don't need to do it
604 if ((*flags
) & LDLM_FL_LOCK_CHANGED
) {
605 int newmode
= reply
->lock_desc
.l_req_mode
;
607 if (newmode
&& newmode
!= lock
->l_req_mode
) {
608 LDLM_DEBUG(lock
, "server returned different mode %s",
609 ldlm_lockname
[newmode
]);
610 lock
->l_req_mode
= newmode
;
613 if (memcmp(reply
->lock_desc
.l_resource
.lr_name
.name
,
614 lock
->l_resource
->lr_name
.name
,
615 sizeof(struct ldlm_res_id
))) {
616 CDEBUG(D_INFO
, "remote intent success, locking "
617 "(%ld,%ld,%ld) instead of "
619 (long)reply
->lock_desc
.l_resource
.lr_name
.name
[0],
620 (long)reply
->lock_desc
.l_resource
.lr_name
.name
[1],
621 (long)reply
->lock_desc
.l_resource
.lr_name
.name
[2],
622 (long)lock
->l_resource
->lr_name
.name
[0],
623 (long)lock
->l_resource
->lr_name
.name
[1],
624 (long)lock
->l_resource
->lr_name
.name
[2]);
626 rc
= ldlm_lock_change_resource(ns
, lock
,
627 &reply
->lock_desc
.l_resource
.lr_name
);
628 if (rc
|| lock
->l_resource
== NULL
)
629 GOTO(cleanup
, rc
= -ENOMEM
);
630 LDLM_DEBUG(lock
, "client-side enqueue, new resource");
633 if (!(type
== LDLM_IBITS
&&
634 !(exp_connect_flags(exp
) & OBD_CONNECT_IBITS
)))
635 /* We assume lock type cannot change on server*/
636 ldlm_convert_policy_to_local(exp
,
637 lock
->l_resource
->lr_type
,
638 &reply
->lock_desc
.l_policy_data
,
639 &lock
->l_policy_data
);
640 if (type
!= LDLM_PLAIN
)
641 LDLM_DEBUG(lock
,"client-side enqueue, new policy data");
644 if ((*flags
) & LDLM_FL_AST_SENT
||
645 /* Cancel extent locks as soon as possible on a liblustre client,
646 * because it cannot handle asynchronous ASTs robustly (see
648 (LIBLUSTRE_CLIENT
&& type
== LDLM_EXTENT
)) {
649 lock_res_and_lock(lock
);
650 lock
->l_flags
|= LDLM_FL_CBPENDING
| LDLM_FL_BL_AST
;
651 unlock_res_and_lock(lock
);
652 LDLM_DEBUG(lock
, "enqueue reply includes blocking AST");
655 /* If the lock has already been granted by a completion AST, don't
656 * clobber the LVB with an older one. */
658 /* We must lock or a racing completion might update lvb without
659 * letting us know and we'll clobber the correct value.
660 * Cannot unlock after the check either, a that still leaves
661 * a tiny window for completion to get in */
662 lock_res_and_lock(lock
);
663 if (lock
->l_req_mode
!= lock
->l_granted_mode
)
664 rc
= ldlm_fill_lvb(lock
, &req
->rq_pill
, RCL_SERVER
,
665 lock
->l_lvb_data
, size
);
666 unlock_res_and_lock(lock
);
674 rc
= ldlm_lock_enqueue(ns
, &lock
, NULL
, flags
);
675 if (lock
->l_completion_ast
!= NULL
) {
676 int err
= lock
->l_completion_ast(lock
, *flags
, NULL
);
684 if (lvb_len
&& lvb
!= NULL
) {
685 /* Copy the LVB here, and not earlier, because the completion
686 * AST (if any) can override what we got in the reply */
687 memcpy(lvb
, lock
->l_lvb_data
, lvb_len
);
690 LDLM_DEBUG(lock
, "client-side enqueue END");
692 if (cleanup_phase
== 1 && rc
)
693 failed_lock_cleanup(ns
, lock
, mode
);
694 /* Put lock 2 times, the second reference is held by ldlm_cli_enqueue */
696 LDLM_LOCK_RELEASE(lock
);
699 EXPORT_SYMBOL(ldlm_cli_enqueue_fini
);
702 * Estimate number of lock handles that would fit into request of given
703 * size. PAGE_SIZE-512 is to allow TCP/IP and LNET headers to fit into
704 * a single page on the send/receive side. XXX: 512 should be changed to
705 * more adequate value.
707 static inline int ldlm_req_handles_avail(int req_size
, int off
)
711 avail
= min_t(int, LDLM_MAXREQSIZE
, PAGE_CACHE_SIZE
- 512) - req_size
;
712 if (likely(avail
>= 0))
713 avail
/= (int)sizeof(struct lustre_handle
);
716 avail
+= LDLM_LOCKREQ_HANDLES
- off
;
721 static inline int ldlm_capsule_handles_avail(struct req_capsule
*pill
,
722 enum req_location loc
,
725 int size
= req_capsule_msg_size(pill
, loc
);
726 return ldlm_req_handles_avail(size
, off
);
729 static inline int ldlm_format_handles_avail(struct obd_import
*imp
,
730 const struct req_format
*fmt
,
731 enum req_location loc
, int off
)
733 int size
= req_capsule_fmt_size(imp
->imp_msg_magic
, fmt
, loc
);
734 return ldlm_req_handles_avail(size
, off
);
738 * Cancel LRU locks and pack them into the enqueue request. Pack there the given
739 * \a count locks in \a cancels.
741 * This is to be called by functions preparing their own requests that
742 * might contain lists of locks to cancel in addition to actual operation
743 * that needs to be performed.
745 int ldlm_prep_elc_req(struct obd_export
*exp
, struct ptlrpc_request
*req
,
746 int version
, int opc
, int canceloff
,
747 struct list_head
*cancels
, int count
)
749 struct ldlm_namespace
*ns
= exp
->exp_obd
->obd_namespace
;
750 struct req_capsule
*pill
= &req
->rq_pill
;
751 struct ldlm_request
*dlm
= NULL
;
752 int flags
, avail
, to_free
, pack
= 0;
758 if (ns_connect_cancelset(ns
)) {
759 /* Estimate the amount of available space in the request. */
760 req_capsule_filled_sizes(pill
, RCL_CLIENT
);
761 avail
= ldlm_capsule_handles_avail(pill
, RCL_CLIENT
, canceloff
);
763 flags
= ns_connect_lru_resize(ns
) ?
764 LDLM_CANCEL_LRUR
: LDLM_CANCEL_AGED
;
765 to_free
= !ns_connect_lru_resize(ns
) &&
766 opc
== LDLM_ENQUEUE
? 1 : 0;
768 /* Cancel LRU locks here _only_ if the server supports
769 * EARLY_CANCEL. Otherwise we have to send extra CANCEL
770 * RPC, which will make us slower. */
772 count
+= ldlm_cancel_lru_local(ns
, cancels
, to_free
,
773 avail
- count
, 0, flags
);
778 req_capsule_set_size(pill
, &RMF_DLM_REQ
, RCL_CLIENT
,
779 ldlm_request_bufsize(pack
, opc
));
782 rc
= ptlrpc_request_pack(req
, version
, opc
);
784 ldlm_lock_list_put(cancels
, l_bl_ast
, count
);
788 if (ns_connect_cancelset(ns
)) {
790 dlm
= req_capsule_client_get(pill
, &RMF_DLM_REQ
);
792 /* Skip first lock handler in ldlm_request_pack(),
793 * this method will incrment @lock_count according
794 * to the lock handle amount actually written to
796 dlm
->lock_count
= canceloff
;
798 /* Pack into the request @pack lock handles. */
799 ldlm_cli_cancel_list(cancels
, pack
, req
, 0);
800 /* Prepare and send separate cancel RPC for others. */
801 ldlm_cli_cancel_list(cancels
, count
- pack
, NULL
, 0);
803 ldlm_lock_list_put(cancels
, l_bl_ast
, count
);
807 EXPORT_SYMBOL(ldlm_prep_elc_req
);
809 int ldlm_prep_enqueue_req(struct obd_export
*exp
, struct ptlrpc_request
*req
,
810 struct list_head
*cancels
, int count
)
812 return ldlm_prep_elc_req(exp
, req
, LUSTRE_DLM_VERSION
, LDLM_ENQUEUE
,
813 LDLM_ENQUEUE_CANCEL_OFF
, cancels
, count
);
815 EXPORT_SYMBOL(ldlm_prep_enqueue_req
);
817 struct ptlrpc_request
*ldlm_enqueue_pack(struct obd_export
*exp
, int lvb_len
)
819 struct ptlrpc_request
*req
;
822 req
= ptlrpc_request_alloc(class_exp2cliimp(exp
), &RQF_LDLM_ENQUEUE
);
824 return ERR_PTR(-ENOMEM
);
826 rc
= ldlm_prep_enqueue_req(exp
, req
, NULL
, 0);
828 ptlrpc_request_free(req
);
832 req_capsule_set_size(&req
->rq_pill
, &RMF_DLM_LVB
, RCL_SERVER
, lvb_len
);
833 ptlrpc_request_set_replen(req
);
836 EXPORT_SYMBOL(ldlm_enqueue_pack
);
839 * Client-side lock enqueue.
841 * If a request has some specific initialisation it is passed in \a reqp,
842 * otherwise it is created in ldlm_cli_enqueue.
844 * Supports sync and async requests, pass \a async flag accordingly. If a
845 * request was created in ldlm_cli_enqueue and it is the async request,
846 * pass it to the caller in \a reqp.
848 int ldlm_cli_enqueue(struct obd_export
*exp
, struct ptlrpc_request
**reqp
,
849 struct ldlm_enqueue_info
*einfo
,
850 const struct ldlm_res_id
*res_id
,
851 ldlm_policy_data_t
const *policy
, __u64
*flags
,
852 void *lvb
, __u32 lvb_len
, enum lvb_type lvb_type
,
853 struct lustre_handle
*lockh
, int async
)
855 struct ldlm_namespace
*ns
;
856 struct ldlm_lock
*lock
;
857 struct ldlm_request
*body
;
858 int is_replay
= *flags
& LDLM_FL_REPLAY
;
859 int req_passed_in
= 1;
861 struct ptlrpc_request
*req
;
863 LASSERT(exp
!= NULL
);
865 ns
= exp
->exp_obd
->obd_namespace
;
867 /* If we're replaying this lock, just check some invariants.
868 * If we're creating a new lock, get everything all setup nice. */
870 lock
= ldlm_handle2lock_long(lockh
, 0);
871 LASSERT(lock
!= NULL
);
872 LDLM_DEBUG(lock
, "client-side enqueue START");
873 LASSERT(exp
== lock
->l_conn_export
);
875 const struct ldlm_callback_suite cbs
= {
876 .lcs_completion
= einfo
->ei_cb_cp
,
877 .lcs_blocking
= einfo
->ei_cb_bl
,
878 .lcs_glimpse
= einfo
->ei_cb_gl
880 lock
= ldlm_lock_create(ns
, res_id
, einfo
->ei_type
,
881 einfo
->ei_mode
, &cbs
, einfo
->ei_cbdata
,
885 /* for the local lock, add the reference */
886 ldlm_lock_addref_internal(lock
, einfo
->ei_mode
);
887 ldlm_lock2handle(lock
, lockh
);
888 if (policy
!= NULL
) {
889 /* INODEBITS_INTEROP: If the server does not support
890 * inodebits, we will request a plain lock in the
891 * descriptor (ldlm_lock2desc() below) but use an
892 * inodebits lock internally with both bits set.
894 if (einfo
->ei_type
== LDLM_IBITS
&&
895 !(exp_connect_flags(exp
) &
897 lock
->l_policy_data
.l_inodebits
.bits
=
898 MDS_INODELOCK_LOOKUP
|
899 MDS_INODELOCK_UPDATE
;
901 lock
->l_policy_data
= *policy
;
904 if (einfo
->ei_type
== LDLM_EXTENT
)
905 lock
->l_req_extent
= policy
->l_extent
;
906 LDLM_DEBUG(lock
, "client-side enqueue START, flags %llx\n",
910 lock
->l_conn_export
= exp
;
911 lock
->l_export
= NULL
;
912 lock
->l_blocking_ast
= einfo
->ei_cb_bl
;
913 lock
->l_flags
|= (*flags
& LDLM_FL_NO_LRU
);
915 /* lock not sent to server yet */
917 if (reqp
== NULL
|| *reqp
== NULL
) {
918 req
= ptlrpc_request_alloc_pack(class_exp2cliimp(exp
),
923 failed_lock_cleanup(ns
, lock
, einfo
->ei_mode
);
924 LDLM_LOCK_RELEASE(lock
);
934 len
= req_capsule_get_size(&req
->rq_pill
, &RMF_DLM_REQ
,
936 LASSERTF(len
>= sizeof(*body
), "buflen[%d] = %d, not %d\n",
937 DLM_LOCKREQ_OFF
, len
, (int)sizeof(*body
));
940 /* Dump lock data into the request buffer */
941 body
= req_capsule_client_get(&req
->rq_pill
, &RMF_DLM_REQ
);
942 ldlm_lock2desc(lock
, &body
->lock_desc
);
943 body
->lock_flags
= ldlm_flags_to_wire(*flags
);
944 body
->lock_handle
[0] = *lockh
;
946 /* Continue as normal. */
947 if (!req_passed_in
) {
949 req_capsule_extend(&req
->rq_pill
,
950 &RQF_LDLM_ENQUEUE_LVB
);
951 req_capsule_set_size(&req
->rq_pill
, &RMF_DLM_LVB
, RCL_SERVER
,
953 ptlrpc_request_set_replen(req
);
957 * Liblustre client doesn't get extent locks, except for O_APPEND case
958 * where [0, OBD_OBJECT_EOF] lock is taken, or truncate, where
959 * [i_size, OBD_OBJECT_EOF] lock is taken.
961 LASSERT(ergo(LIBLUSTRE_CLIENT
, einfo
->ei_type
!= LDLM_EXTENT
||
962 policy
->l_extent
.end
== OBD_OBJECT_EOF
));
965 LASSERT(reqp
!= NULL
);
969 LDLM_DEBUG(lock
, "sending request");
971 rc
= ptlrpc_queue_wait(req
);
973 err
= ldlm_cli_enqueue_fini(exp
, req
, einfo
->ei_type
, policy
? 1 : 0,
974 einfo
->ei_mode
, flags
, lvb
, lvb_len
,
977 /* If ldlm_cli_enqueue_fini did not find the lock, we need to free
978 * one reference that we took */
980 LDLM_LOCK_RELEASE(lock
);
984 if (!req_passed_in
&& req
!= NULL
) {
985 ptlrpc_req_finished(req
);
992 EXPORT_SYMBOL(ldlm_cli_enqueue
);
994 static int ldlm_cli_convert_local(struct ldlm_lock
*lock
, int new_mode
,
997 struct ldlm_resource
*res
;
1000 if (ns_is_client(ldlm_lock_to_ns(lock
))) {
1001 CERROR("Trying to cancel local lock\n");
1004 LDLM_DEBUG(lock
, "client-side local convert");
1006 res
= ldlm_lock_convert(lock
, new_mode
, flags
);
1008 ldlm_reprocess_all(res
);
1011 rc
= LUSTRE_EDEADLK
;
1013 LDLM_DEBUG(lock
, "client-side local convert handler END");
1014 LDLM_LOCK_PUT(lock
);
1018 /* FIXME: one of ldlm_cli_convert or the server side should reject attempted
1019 * conversion of locks which are on the waiting or converting queue */
1020 /* Caller of this code is supposed to take care of lock readers/writers
1022 int ldlm_cli_convert(struct lustre_handle
*lockh
, int new_mode
, __u32
*flags
)
1024 struct ldlm_request
*body
;
1025 struct ldlm_reply
*reply
;
1026 struct ldlm_lock
*lock
;
1027 struct ldlm_resource
*res
;
1028 struct ptlrpc_request
*req
;
1031 lock
= ldlm_handle2lock(lockh
);
1038 if (lock
->l_conn_export
== NULL
)
1039 return ldlm_cli_convert_local(lock
, new_mode
, flags
);
1041 LDLM_DEBUG(lock
, "client-side convert");
1043 req
= ptlrpc_request_alloc_pack(class_exp2cliimp(lock
->l_conn_export
),
1044 &RQF_LDLM_CONVERT
, LUSTRE_DLM_VERSION
,
1047 LDLM_LOCK_PUT(lock
);
1051 body
= req_capsule_client_get(&req
->rq_pill
, &RMF_DLM_REQ
);
1052 body
->lock_handle
[0] = lock
->l_remote_handle
;
1054 body
->lock_desc
.l_req_mode
= new_mode
;
1055 body
->lock_flags
= ldlm_flags_to_wire(*flags
);
1058 ptlrpc_request_set_replen(req
);
1059 rc
= ptlrpc_queue_wait(req
);
1063 reply
= req_capsule_server_get(&req
->rq_pill
, &RMF_DLM_REP
);
1065 GOTO(out
, rc
= -EPROTO
);
1068 GOTO(out
, rc
= req
->rq_status
);
1070 res
= ldlm_lock_convert(lock
, new_mode
, &reply
->lock_flags
);
1072 ldlm_reprocess_all(res
);
1073 /* Go to sleep until the lock is granted. */
1074 /* FIXME: or cancelled. */
1075 if (lock
->l_completion_ast
) {
1076 rc
= lock
->l_completion_ast(lock
, LDLM_FL_WAIT_NOREPROC
,
1082 rc
= LUSTRE_EDEADLK
;
1085 LDLM_LOCK_PUT(lock
);
1086 ptlrpc_req_finished(req
);
1089 EXPORT_SYMBOL(ldlm_cli_convert
);
1092 * Cancel locks locally.
1094 * \retval LDLM_FL_LOCAL_ONLY if there is no need for a CANCEL RPC to the server
1095 * \retval LDLM_FL_CANCELING otherwise;
1096 * \retval LDLM_FL_BL_AST if there is a need for a separate CANCEL RPC.
1098 static __u64
ldlm_cli_cancel_local(struct ldlm_lock
*lock
)
1100 __u64 rc
= LDLM_FL_LOCAL_ONLY
;
1102 if (lock
->l_conn_export
) {
1105 LDLM_DEBUG(lock
, "client-side cancel");
1106 /* Set this flag to prevent others from getting new references*/
1107 lock_res_and_lock(lock
);
1108 lock
->l_flags
|= LDLM_FL_CBPENDING
;
1109 local_only
= !!(lock
->l_flags
&
1110 (LDLM_FL_LOCAL_ONLY
|LDLM_FL_CANCEL_ON_BLOCK
));
1111 ldlm_cancel_callback(lock
);
1112 rc
= (lock
->l_flags
& LDLM_FL_BL_AST
) ?
1113 LDLM_FL_BL_AST
: LDLM_FL_CANCELING
;
1114 unlock_res_and_lock(lock
);
1117 CDEBUG(D_DLMTRACE
, "not sending request (at caller's "
1119 rc
= LDLM_FL_LOCAL_ONLY
;
1121 ldlm_lock_cancel(lock
);
1123 if (ns_is_client(ldlm_lock_to_ns(lock
))) {
1124 LDLM_ERROR(lock
, "Trying to cancel local lock");
1127 LDLM_DEBUG(lock
, "server-side local cancel");
1128 ldlm_lock_cancel(lock
);
1129 ldlm_reprocess_all(lock
->l_resource
);
1136 * Pack \a count locks in \a head into ldlm_request buffer of request \a req.
1138 static void ldlm_cancel_pack(struct ptlrpc_request
*req
,
1139 struct list_head
*head
, int count
)
1141 struct ldlm_request
*dlm
;
1142 struct ldlm_lock
*lock
;
1143 int max
, packed
= 0;
1145 dlm
= req_capsule_client_get(&req
->rq_pill
, &RMF_DLM_REQ
);
1146 LASSERT(dlm
!= NULL
);
1148 /* Check the room in the request buffer. */
1149 max
= req_capsule_get_size(&req
->rq_pill
, &RMF_DLM_REQ
, RCL_CLIENT
) -
1150 sizeof(struct ldlm_request
);
1151 max
/= sizeof(struct lustre_handle
);
1152 max
+= LDLM_LOCKREQ_HANDLES
;
1153 LASSERT(max
>= dlm
->lock_count
+ count
);
1155 /* XXX: it would be better to pack lock handles grouped by resource.
1156 * so that the server cancel would call filter_lvbo_update() less
1158 list_for_each_entry(lock
, head
, l_bl_ast
) {
1161 LASSERT(lock
->l_conn_export
);
1162 /* Pack the lock handle to the given request buffer. */
1163 LDLM_DEBUG(lock
, "packing");
1164 dlm
->lock_handle
[dlm
->lock_count
++] = lock
->l_remote_handle
;
1167 CDEBUG(D_DLMTRACE
, "%d locks packed\n", packed
);
1171 * Prepare and send a batched cancel RPC. It will include \a count lock
1172 * handles of locks given in \a cancels list. */
1173 int ldlm_cli_cancel_req(struct obd_export
*exp
, struct list_head
*cancels
,
1174 int count
, ldlm_cancel_flags_t flags
)
1176 struct ptlrpc_request
*req
= NULL
;
1177 struct obd_import
*imp
;
1181 LASSERT(exp
!= NULL
);
1184 CFS_FAIL_TIMEOUT(OBD_FAIL_LDLM_PAUSE_CANCEL
, cfs_fail_val
);
1186 if (CFS_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_RACE
))
1189 free
= ldlm_format_handles_avail(class_exp2cliimp(exp
),
1190 &RQF_LDLM_CANCEL
, RCL_CLIENT
, 0);
1195 imp
= class_exp2cliimp(exp
);
1196 if (imp
== NULL
|| imp
->imp_invalid
) {
1198 "skipping cancel on invalid import %p\n", imp
);
1202 req
= ptlrpc_request_alloc(imp
, &RQF_LDLM_CANCEL
);
1204 GOTO(out
, rc
= -ENOMEM
);
1206 req_capsule_filled_sizes(&req
->rq_pill
, RCL_CLIENT
);
1207 req_capsule_set_size(&req
->rq_pill
, &RMF_DLM_REQ
, RCL_CLIENT
,
1208 ldlm_request_bufsize(count
, LDLM_CANCEL
));
1210 rc
= ptlrpc_request_pack(req
, LUSTRE_DLM_VERSION
, LDLM_CANCEL
);
1212 ptlrpc_request_free(req
);
1216 req
->rq_request_portal
= LDLM_CANCEL_REQUEST_PORTAL
;
1217 req
->rq_reply_portal
= LDLM_CANCEL_REPLY_PORTAL
;
1218 ptlrpc_at_set_req_timeout(req
);
1220 ldlm_cancel_pack(req
, cancels
, count
);
1222 ptlrpc_request_set_replen(req
);
1223 if (flags
& LCF_ASYNC
) {
1224 ptlrpcd_add_req(req
, PDL_POLICY_LOCAL
, -1);
1228 rc
= ptlrpc_queue_wait(req
);
1230 if (rc
== LUSTRE_ESTALE
) {
1231 CDEBUG(D_DLMTRACE
, "client/server (nid %s) "
1232 "out of sync -- not fatal\n",
1233 libcfs_nid2str(req
->rq_import
->
1234 imp_connection
->c_peer
.nid
));
1236 } else if (rc
== -ETIMEDOUT
&& /* check there was no reconnect*/
1237 req
->rq_import_generation
== imp
->imp_generation
) {
1238 ptlrpc_req_finished(req
);
1240 } else if (rc
!= ELDLM_OK
) {
1241 /* -ESHUTDOWN is common on umount */
1242 CDEBUG_LIMIT(rc
== -ESHUTDOWN
? D_DLMTRACE
: D_ERROR
,
1243 "Got rc %d from cancel RPC: "
1244 "canceling anyway\n", rc
);
1251 ptlrpc_req_finished(req
);
1253 return sent
? sent
: rc
;
1255 EXPORT_SYMBOL(ldlm_cli_cancel_req
);
1257 static inline struct ldlm_pool
*ldlm_imp2pl(struct obd_import
*imp
)
1259 LASSERT(imp
!= NULL
);
1260 return &imp
->imp_obd
->obd_namespace
->ns_pool
;
1264 * Update client's OBD pool related fields with new SLV and Limit from \a req.
1266 int ldlm_cli_update_pool(struct ptlrpc_request
*req
)
1268 struct obd_device
*obd
;
1272 if (unlikely(!req
->rq_import
|| !req
->rq_import
->imp_obd
||
1273 !imp_connect_lru_resize(req
->rq_import
)))
1276 * Do nothing for corner cases.
1281 /* In some cases RPC may contain SLV and limit zeroed out. This
1282 * is the case when server does not support LRU resize feature.
1283 * This is also possible in some recovery cases when server-side
1284 * reqs have no reference to the OBD export and thus access to
1285 * server-side namespace is not possible. */
1286 if (lustre_msg_get_slv(req
->rq_repmsg
) == 0 ||
1287 lustre_msg_get_limit(req
->rq_repmsg
) == 0) {
1288 DEBUG_REQ(D_HA
, req
, "Zero SLV or Limit found "
1289 "(SLV: "LPU64
", Limit: %u)",
1290 lustre_msg_get_slv(req
->rq_repmsg
),
1291 lustre_msg_get_limit(req
->rq_repmsg
));
1295 new_limit
= lustre_msg_get_limit(req
->rq_repmsg
);
1296 new_slv
= lustre_msg_get_slv(req
->rq_repmsg
);
1297 obd
= req
->rq_import
->imp_obd
;
1299 /* Set new SLV and limit in OBD fields to make them accessible
1300 * to the pool thread. We do not access obd_namespace and pool
1301 * directly here as there is no reliable way to make sure that
1302 * they are still alive at cleanup time. Evil races are possible
1303 * which may cause Oops at that time. */
1304 write_lock(&obd
->obd_pool_lock
);
1305 obd
->obd_pool_slv
= new_slv
;
1306 obd
->obd_pool_limit
= new_limit
;
1307 write_unlock(&obd
->obd_pool_lock
);
1311 EXPORT_SYMBOL(ldlm_cli_update_pool
);
1314 * Client side lock cancel.
1316 * Lock must not have any readers or writers by this time.
1318 int ldlm_cli_cancel(struct lustre_handle
*lockh
,
1319 ldlm_cancel_flags_t cancel_flags
)
1321 struct obd_export
*exp
;
1322 int avail
, flags
, count
= 1;
1324 struct ldlm_namespace
*ns
;
1325 struct ldlm_lock
*lock
;
1328 /* concurrent cancels on the same handle can happen */
1329 lock
= ldlm_handle2lock_long(lockh
, LDLM_FL_CANCELING
);
1331 LDLM_DEBUG_NOLOCK("lock is already being destroyed\n");
1335 rc
= ldlm_cli_cancel_local(lock
);
1336 if (rc
== LDLM_FL_LOCAL_ONLY
) {
1337 LDLM_LOCK_RELEASE(lock
);
1340 /* Even if the lock is marked as LDLM_FL_BL_AST, this is a LDLM_CANCEL
1341 * RPC which goes to canceld portal, so we can cancel other LRU locks
1342 * here and send them all as one LDLM_CANCEL RPC. */
1343 LASSERT(list_empty(&lock
->l_bl_ast
));
1344 list_add(&lock
->l_bl_ast
, &cancels
);
1346 exp
= lock
->l_conn_export
;
1347 if (exp_connect_cancelset(exp
)) {
1348 avail
= ldlm_format_handles_avail(class_exp2cliimp(exp
),
1353 ns
= ldlm_lock_to_ns(lock
);
1354 flags
= ns_connect_lru_resize(ns
) ?
1355 LDLM_CANCEL_LRUR
: LDLM_CANCEL_AGED
;
1356 count
+= ldlm_cancel_lru_local(ns
, &cancels
, 0, avail
- 1,
1359 ldlm_cli_cancel_list(&cancels
, count
, NULL
, cancel_flags
);
1362 EXPORT_SYMBOL(ldlm_cli_cancel
);
1365 * Locally cancel up to \a count locks in list \a cancels.
1366 * Return the number of cancelled locks.
1368 int ldlm_cli_cancel_list_local(struct list_head
*cancels
, int count
,
1369 ldlm_cancel_flags_t flags
)
1372 struct ldlm_lock
*lock
, *next
;
1373 int left
= 0, bl_ast
= 0;
1377 list_for_each_entry_safe(lock
, next
, cancels
, l_bl_ast
) {
1381 if (flags
& LCF_LOCAL
) {
1382 rc
= LDLM_FL_LOCAL_ONLY
;
1383 ldlm_lock_cancel(lock
);
1385 rc
= ldlm_cli_cancel_local(lock
);
1387 /* Until we have compound requests and can send LDLM_CANCEL
1388 * requests batched with generic RPCs, we need to send cancels
1389 * with the LDLM_FL_BL_AST flag in a separate RPC from
1390 * the one being generated now. */
1391 if (!(flags
& LCF_BL_AST
) && (rc
== LDLM_FL_BL_AST
)) {
1392 LDLM_DEBUG(lock
, "Cancel lock separately");
1393 list_del_init(&lock
->l_bl_ast
);
1394 list_add(&lock
->l_bl_ast
, &head
);
1398 if (rc
== LDLM_FL_LOCAL_ONLY
) {
1399 /* CANCEL RPC should not be sent to server. */
1400 list_del_init(&lock
->l_bl_ast
);
1401 LDLM_LOCK_RELEASE(lock
);
1407 ldlm_cli_cancel_list(&head
, bl_ast
, NULL
, 0);
1412 EXPORT_SYMBOL(ldlm_cli_cancel_list_local
);
1415 * Cancel as many locks as possible w/o sending any RPCs (e.g. to write back
1416 * dirty data, to close a file, ...) or waiting for any RPCs in-flight (e.g.
1417 * readahead requests, ...)
1419 static ldlm_policy_res_t
ldlm_cancel_no_wait_policy(struct ldlm_namespace
*ns
,
1420 struct ldlm_lock
*lock
,
1421 int unused
, int added
,
1424 ldlm_policy_res_t result
= LDLM_POLICY_CANCEL_LOCK
;
1425 ldlm_cancel_for_recovery cb
= ns
->ns_cancel_for_recovery
;
1426 lock_res_and_lock(lock
);
1428 /* don't check added & count since we want to process all locks
1429 * from unused list */
1430 switch (lock
->l_resource
->lr_type
) {
1436 result
= LDLM_POLICY_SKIP_LOCK
;
1437 lock
->l_flags
|= LDLM_FL_SKIPPED
;
1441 unlock_res_and_lock(lock
);
1446 * Callback function for LRU-resize policy. Decides whether to keep
1447 * \a lock in LRU for current \a LRU size \a unused, added in current
1448 * scan \a added and number of locks to be preferably canceled \a count.
1450 * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1452 * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1454 static ldlm_policy_res_t
ldlm_cancel_lrur_policy(struct ldlm_namespace
*ns
,
1455 struct ldlm_lock
*lock
,
1456 int unused
, int added
,
1459 cfs_time_t cur
= cfs_time_current();
1460 struct ldlm_pool
*pl
= &ns
->ns_pool
;
1464 /* Stop LRU processing when we reach past @count or have checked all
1466 if (count
&& added
>= count
)
1467 return LDLM_POLICY_KEEP_LOCK
;
1469 slv
= ldlm_pool_get_slv(pl
);
1470 lvf
= ldlm_pool_get_lvf(pl
);
1471 la
= cfs_duration_sec(cfs_time_sub(cur
,
1472 lock
->l_last_used
));
1473 lv
= lvf
* la
* unused
;
1475 /* Inform pool about current CLV to see it via proc. */
1476 ldlm_pool_set_clv(pl
, lv
);
1478 /* Stop when SLV is not yet come from server or lv is smaller than
1480 return (slv
== 0 || lv
< slv
) ?
1481 LDLM_POLICY_KEEP_LOCK
: LDLM_POLICY_CANCEL_LOCK
;
1485 * Callback function for proc used policy. Makes decision whether to keep
1486 * \a lock in LRU for current \a LRU size \a unused, added in current scan \a
1487 * added and number of locks to be preferably canceled \a count.
1489 * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1491 * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1493 static ldlm_policy_res_t
ldlm_cancel_passed_policy(struct ldlm_namespace
*ns
,
1494 struct ldlm_lock
*lock
,
1495 int unused
, int added
,
1498 /* Stop LRU processing when we reach past @count or have checked all
1500 return (added
>= count
) ?
1501 LDLM_POLICY_KEEP_LOCK
: LDLM_POLICY_CANCEL_LOCK
;
1505 * Callback function for aged policy. Makes decision whether to keep \a lock in
1506 * LRU for current LRU size \a unused, added in current scan \a added and
1507 * number of locks to be preferably canceled \a count.
1509 * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1511 * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1513 static ldlm_policy_res_t
ldlm_cancel_aged_policy(struct ldlm_namespace
*ns
,
1514 struct ldlm_lock
*lock
,
1515 int unused
, int added
,
1518 /* Stop LRU processing if young lock is found and we reach past count */
1519 return ((added
>= count
) &&
1520 cfs_time_before(cfs_time_current(),
1521 cfs_time_add(lock
->l_last_used
,
1523 LDLM_POLICY_KEEP_LOCK
: LDLM_POLICY_CANCEL_LOCK
;
1527 * Callback function for default policy. Makes decision whether to keep \a lock
1528 * in LRU for current LRU size \a unused, added in current scan \a added and
1529 * number of locks to be preferably canceled \a count.
1531 * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1533 * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1535 static ldlm_policy_res_t
ldlm_cancel_default_policy(struct ldlm_namespace
*ns
,
1536 struct ldlm_lock
*lock
,
1537 int unused
, int added
,
1540 /* Stop LRU processing when we reach past count or have checked all
1542 return (added
>= count
) ?
1543 LDLM_POLICY_KEEP_LOCK
: LDLM_POLICY_CANCEL_LOCK
;
1546 typedef ldlm_policy_res_t (*ldlm_cancel_lru_policy_t
)(struct ldlm_namespace
*,
1547 struct ldlm_lock
*, int,
1550 static ldlm_cancel_lru_policy_t
1551 ldlm_cancel_lru_policy(struct ldlm_namespace
*ns
, int flags
)
1553 if (flags
& LDLM_CANCEL_NO_WAIT
)
1554 return ldlm_cancel_no_wait_policy
;
1556 if (ns_connect_lru_resize(ns
)) {
1557 if (flags
& LDLM_CANCEL_SHRINK
)
1558 /* We kill passed number of old locks. */
1559 return ldlm_cancel_passed_policy
;
1560 else if (flags
& LDLM_CANCEL_LRUR
)
1561 return ldlm_cancel_lrur_policy
;
1562 else if (flags
& LDLM_CANCEL_PASSED
)
1563 return ldlm_cancel_passed_policy
;
1565 if (flags
& LDLM_CANCEL_AGED
)
1566 return ldlm_cancel_aged_policy
;
1569 return ldlm_cancel_default_policy
;
1573 * - Free space in LRU for \a count new locks,
1574 * redundant unused locks are canceled locally;
1575 * - also cancel locally unused aged locks;
1576 * - do not cancel more than \a max locks;
1577 * - GET the found locks and add them into the \a cancels list.
1579 * A client lock can be added to the l_bl_ast list only when it is
1580 * marked LDLM_FL_CANCELING. Otherwise, somebody is already doing
1581 * CANCEL. There are the following use cases:
1582 * ldlm_cancel_resource_local(), ldlm_cancel_lru_local() and
1583 * ldlm_cli_cancel(), which check and set this flag properly. As any
1584 * attempt to cancel a lock rely on this flag, l_bl_ast list is accessed
1585 * later without any special locking.
1587 * Calling policies for enabled LRU resize:
1588 * ----------------------------------------
1589 * flags & LDLM_CANCEL_LRUR - use LRU resize policy (SLV from server) to
1590 * cancel not more than \a count locks;
1592 * flags & LDLM_CANCEL_PASSED - cancel \a count number of old locks (located at
1593 * the beginning of LRU list);
1595 * flags & LDLM_CANCEL_SHRINK - cancel not more than \a count locks according to
1596 * memory pressre policy function;
1598 * flags & LDLM_CANCEL_AGED - cancel \a count locks according to "aged policy".
1600 * flags & LDLM_CANCEL_NO_WAIT - cancel as many unused locks as possible
1601 * (typically before replaying locks) w/o
1602 * sending any RPCs or waiting for any
1603 * outstanding RPC to complete.
1605 static int ldlm_prepare_lru_list(struct ldlm_namespace
*ns
, struct list_head
*cancels
,
1606 int count
, int max
, int flags
)
1608 ldlm_cancel_lru_policy_t pf
;
1609 struct ldlm_lock
*lock
, *next
;
1610 int added
= 0, unused
, remained
;
1612 spin_lock(&ns
->ns_lock
);
1613 unused
= ns
->ns_nr_unused
;
1616 if (!ns_connect_lru_resize(ns
))
1617 count
+= unused
- ns
->ns_max_unused
;
1619 pf
= ldlm_cancel_lru_policy(ns
, flags
);
1620 LASSERT(pf
!= NULL
);
1622 while (!list_empty(&ns
->ns_unused_list
)) {
1623 ldlm_policy_res_t result
;
1625 /* all unused locks */
1626 if (remained
-- <= 0)
1629 /* For any flags, stop scanning if @max is reached. */
1630 if (max
&& added
>= max
)
1633 list_for_each_entry_safe(lock
, next
, &ns
->ns_unused_list
,
1635 /* No locks which got blocking requests. */
1636 LASSERT(!(lock
->l_flags
& LDLM_FL_BL_AST
));
1638 if (flags
& LDLM_CANCEL_NO_WAIT
&&
1639 lock
->l_flags
& LDLM_FL_SKIPPED
)
1640 /* already processed */
1643 /* Somebody is already doing CANCEL. No need for this
1644 * lock in LRU, do not traverse it again. */
1645 if (!(lock
->l_flags
& LDLM_FL_CANCELING
))
1648 ldlm_lock_remove_from_lru_nolock(lock
);
1650 if (&lock
->l_lru
== &ns
->ns_unused_list
)
1653 LDLM_LOCK_GET(lock
);
1654 spin_unlock(&ns
->ns_lock
);
1655 lu_ref_add(&lock
->l_reference
, __func__
, current
);
1657 /* Pass the lock through the policy filter and see if it
1658 * should stay in LRU.
1660 * Even for shrinker policy we stop scanning if
1661 * we find a lock that should stay in the cache.
1662 * We should take into account lock age anyway
1663 * as a new lock is a valuable resource even if
1664 * it has a low weight.
1666 * That is, for shrinker policy we drop only
1667 * old locks, but additionally choose them by
1668 * their weight. Big extent locks will stay in
1670 result
= pf(ns
, lock
, unused
, added
, count
);
1671 if (result
== LDLM_POLICY_KEEP_LOCK
) {
1672 lu_ref_del(&lock
->l_reference
,
1674 LDLM_LOCK_RELEASE(lock
);
1675 spin_lock(&ns
->ns_lock
);
1678 if (result
== LDLM_POLICY_SKIP_LOCK
) {
1679 lu_ref_del(&lock
->l_reference
,
1681 LDLM_LOCK_RELEASE(lock
);
1682 spin_lock(&ns
->ns_lock
);
1686 lock_res_and_lock(lock
);
1687 /* Check flags again under the lock. */
1688 if ((lock
->l_flags
& LDLM_FL_CANCELING
) ||
1689 (ldlm_lock_remove_from_lru(lock
) == 0)) {
1690 /* Another thread is removing lock from LRU, or
1691 * somebody is already doing CANCEL, or there
1692 * is a blocking request which will send cancel
1693 * by itself, or the lock is no longer unused. */
1694 unlock_res_and_lock(lock
);
1695 lu_ref_del(&lock
->l_reference
,
1697 LDLM_LOCK_RELEASE(lock
);
1698 spin_lock(&ns
->ns_lock
);
1701 LASSERT(!lock
->l_readers
&& !lock
->l_writers
);
1703 /* If we have chosen to cancel this lock voluntarily, we
1704 * better send cancel notification to server, so that it
1705 * frees appropriate state. This might lead to a race
1706 * where while we are doing cancel here, server is also
1707 * silently cancelling this lock. */
1708 lock
->l_flags
&= ~LDLM_FL_CANCEL_ON_BLOCK
;
1710 /* Setting the CBPENDING flag is a little misleading,
1711 * but prevents an important race; namely, once
1712 * CBPENDING is set, the lock can accumulate no more
1713 * readers/writers. Since readers and writers are
1714 * already zero here, ldlm_lock_decref() won't see
1715 * this flag and call l_blocking_ast */
1716 lock
->l_flags
|= LDLM_FL_CBPENDING
| LDLM_FL_CANCELING
;
1718 /* We can't re-add to l_lru as it confuses the
1719 * refcounting in ldlm_lock_remove_from_lru() if an AST
1720 * arrives after we drop lr_lock below. We use l_bl_ast
1721 * and can't use l_pending_chain as it is used both on
1722 * server and client nevertheless bug 5666 says it is
1723 * used only on server */
1724 LASSERT(list_empty(&lock
->l_bl_ast
));
1725 list_add(&lock
->l_bl_ast
, cancels
);
1726 unlock_res_and_lock(lock
);
1727 lu_ref_del(&lock
->l_reference
, __func__
, current
);
1728 spin_lock(&ns
->ns_lock
);
1732 spin_unlock(&ns
->ns_lock
);
1736 int ldlm_cancel_lru_local(struct ldlm_namespace
*ns
, struct list_head
*cancels
,
1737 int count
, int max
, ldlm_cancel_flags_t cancel_flags
,
1741 added
= ldlm_prepare_lru_list(ns
, cancels
, count
, max
, flags
);
1744 return ldlm_cli_cancel_list_local(cancels
, added
, cancel_flags
);
1748 * Cancel at least \a nr locks from given namespace LRU.
1750 * When called with LCF_ASYNC the blocking callback will be handled
1751 * in a thread and this function will return after the thread has been
1752 * asked to call the callback. When called with LCF_ASYNC the blocking
1753 * callback will be performed in this function.
1755 int ldlm_cancel_lru(struct ldlm_namespace
*ns
, int nr
,
1756 ldlm_cancel_flags_t cancel_flags
,
1762 /* Just prepare the list of locks, do not actually cancel them yet.
1763 * Locks are cancelled later in a separate thread. */
1764 count
= ldlm_prepare_lru_list(ns
, &cancels
, nr
, 0, flags
);
1765 rc
= ldlm_bl_to_thread_list(ns
, NULL
, &cancels
, count
, cancel_flags
);
1773 * Find and cancel locally unused locks found on resource, matched to the
1774 * given policy, mode. GET the found locks and add them into the \a cancels
1777 int ldlm_cancel_resource_local(struct ldlm_resource
*res
,
1778 struct list_head
*cancels
,
1779 ldlm_policy_data_t
*policy
,
1780 ldlm_mode_t mode
, int lock_flags
,
1781 ldlm_cancel_flags_t cancel_flags
, void *opaque
)
1783 struct ldlm_lock
*lock
;
1787 list_for_each_entry(lock
, &res
->lr_granted
, l_res_link
) {
1788 if (opaque
!= NULL
&& lock
->l_ast_data
!= opaque
) {
1789 LDLM_ERROR(lock
, "data %p doesn't match opaque %p",
1790 lock
->l_ast_data
, opaque
);
1795 if (lock
->l_readers
|| lock
->l_writers
)
1798 /* If somebody is already doing CANCEL, or blocking AST came,
1799 * skip this lock. */
1800 if (lock
->l_flags
& LDLM_FL_BL_AST
||
1801 lock
->l_flags
& LDLM_FL_CANCELING
)
1804 if (lockmode_compat(lock
->l_granted_mode
, mode
))
1807 /* If policy is given and this is IBITS lock, add to list only
1808 * those locks that match by policy. */
1809 if (policy
&& (lock
->l_resource
->lr_type
== LDLM_IBITS
) &&
1810 !(lock
->l_policy_data
.l_inodebits
.bits
&
1811 policy
->l_inodebits
.bits
))
1814 /* See CBPENDING comment in ldlm_cancel_lru */
1815 lock
->l_flags
|= LDLM_FL_CBPENDING
| LDLM_FL_CANCELING
|
1818 LASSERT(list_empty(&lock
->l_bl_ast
));
1819 list_add(&lock
->l_bl_ast
, cancels
);
1820 LDLM_LOCK_GET(lock
);
1825 return ldlm_cli_cancel_list_local(cancels
, count
, cancel_flags
);
1827 EXPORT_SYMBOL(ldlm_cancel_resource_local
);
1830 * Cancel client-side locks from a list and send/prepare cancel RPCs to the
1832 * If \a req is NULL, send CANCEL request to server with handles of locks
1833 * in the \a cancels. If EARLY_CANCEL is not supported, send CANCEL requests
1834 * separately per lock.
1835 * If \a req is not NULL, put handles of locks in \a cancels into the request
1836 * buffer at the offset \a off.
1837 * Destroy \a cancels at the end.
1839 int ldlm_cli_cancel_list(struct list_head
*cancels
, int count
,
1840 struct ptlrpc_request
*req
, ldlm_cancel_flags_t flags
)
1842 struct ldlm_lock
*lock
;
1845 if (list_empty(cancels
) || count
== 0)
1848 /* XXX: requests (both batched and not) could be sent in parallel.
1849 * Usually it is enough to have just 1 RPC, but it is possible that
1850 * there are too many locks to be cancelled in LRU or on a resource.
1851 * It would also speed up the case when the server does not support
1854 LASSERT(!list_empty(cancels
));
1855 lock
= list_entry(cancels
->next
, struct ldlm_lock
,
1857 LASSERT(lock
->l_conn_export
);
1859 if (exp_connect_cancelset(lock
->l_conn_export
)) {
1862 ldlm_cancel_pack(req
, cancels
, count
);
1864 res
= ldlm_cli_cancel_req(lock
->l_conn_export
,
1868 res
= ldlm_cli_cancel_req(lock
->l_conn_export
,
1873 CDEBUG_LIMIT(res
== -ESHUTDOWN
? D_DLMTRACE
: D_ERROR
,
1874 "ldlm_cli_cancel_list: %d\n", res
);
1879 ldlm_lock_list_put(cancels
, l_bl_ast
, res
);
1881 LASSERT(count
== 0);
1884 EXPORT_SYMBOL(ldlm_cli_cancel_list
);
1887 * Cancel all locks on a resource that have 0 readers/writers.
1889 * If flags & LDLM_FL_LOCAL_ONLY, throw the locks away without trying
1890 * to notify the server. */
1891 int ldlm_cli_cancel_unused_resource(struct ldlm_namespace
*ns
,
1892 const struct ldlm_res_id
*res_id
,
1893 ldlm_policy_data_t
*policy
,
1895 ldlm_cancel_flags_t flags
,
1898 struct ldlm_resource
*res
;
1903 res
= ldlm_resource_get(ns
, NULL
, res_id
, 0, 0);
1905 /* This is not a problem. */
1906 CDEBUG(D_INFO
, "No resource "LPU64
"\n", res_id
->name
[0]);
1910 LDLM_RESOURCE_ADDREF(res
);
1911 count
= ldlm_cancel_resource_local(res
, &cancels
, policy
, mode
,
1912 0, flags
| LCF_BL_AST
, opaque
);
1913 rc
= ldlm_cli_cancel_list(&cancels
, count
, NULL
, flags
);
1915 CERROR("ldlm_cli_cancel_unused_resource: %d\n", rc
);
1917 LDLM_RESOURCE_DELREF(res
);
1918 ldlm_resource_putref(res
);
1921 EXPORT_SYMBOL(ldlm_cli_cancel_unused_resource
);
1923 struct ldlm_cli_cancel_arg
{
1928 static int ldlm_cli_hash_cancel_unused(struct cfs_hash
*hs
, struct cfs_hash_bd
*bd
,
1929 struct hlist_node
*hnode
, void *arg
)
1931 struct ldlm_resource
*res
= cfs_hash_object(hs
, hnode
);
1932 struct ldlm_cli_cancel_arg
*lc
= arg
;
1935 rc
= ldlm_cli_cancel_unused_resource(ldlm_res_to_ns(res
), &res
->lr_name
,
1937 lc
->lc_flags
, lc
->lc_opaque
);
1939 CERROR("ldlm_cli_cancel_unused ("LPU64
"): %d\n",
1940 res
->lr_name
.name
[0], rc
);
1942 /* must return 0 for hash iteration */
1947 * Cancel all locks on a namespace (or a specific resource, if given)
1948 * that have 0 readers/writers.
1950 * If flags & LCF_LOCAL, throw the locks away without trying
1951 * to notify the server. */
1952 int ldlm_cli_cancel_unused(struct ldlm_namespace
*ns
,
1953 const struct ldlm_res_id
*res_id
,
1954 ldlm_cancel_flags_t flags
, void *opaque
)
1956 struct ldlm_cli_cancel_arg arg
= {
1958 .lc_opaque
= opaque
,
1964 if (res_id
!= NULL
) {
1965 return ldlm_cli_cancel_unused_resource(ns
, res_id
, NULL
,
1969 cfs_hash_for_each_nolock(ns
->ns_rs_hash
,
1970 ldlm_cli_hash_cancel_unused
, &arg
);
1974 EXPORT_SYMBOL(ldlm_cli_cancel_unused
);
1976 /* Lock iterators. */
1978 int ldlm_resource_foreach(struct ldlm_resource
*res
, ldlm_iterator_t iter
,
1981 struct list_head
*tmp
, *next
;
1982 struct ldlm_lock
*lock
;
1983 int rc
= LDLM_ITER_CONTINUE
;
1986 return LDLM_ITER_CONTINUE
;
1989 list_for_each_safe(tmp
, next
, &res
->lr_granted
) {
1990 lock
= list_entry(tmp
, struct ldlm_lock
, l_res_link
);
1992 if (iter(lock
, closure
) == LDLM_ITER_STOP
)
1993 GOTO(out
, rc
= LDLM_ITER_STOP
);
1996 list_for_each_safe(tmp
, next
, &res
->lr_converting
) {
1997 lock
= list_entry(tmp
, struct ldlm_lock
, l_res_link
);
1999 if (iter(lock
, closure
) == LDLM_ITER_STOP
)
2000 GOTO(out
, rc
= LDLM_ITER_STOP
);
2003 list_for_each_safe(tmp
, next
, &res
->lr_waiting
) {
2004 lock
= list_entry(tmp
, struct ldlm_lock
, l_res_link
);
2006 if (iter(lock
, closure
) == LDLM_ITER_STOP
)
2007 GOTO(out
, rc
= LDLM_ITER_STOP
);
2013 EXPORT_SYMBOL(ldlm_resource_foreach
);
2015 struct iter_helper_data
{
2016 ldlm_iterator_t iter
;
2020 static int ldlm_iter_helper(struct ldlm_lock
*lock
, void *closure
)
2022 struct iter_helper_data
*helper
= closure
;
2023 return helper
->iter(lock
, helper
->closure
);
2026 static int ldlm_res_iter_helper(struct cfs_hash
*hs
, struct cfs_hash_bd
*bd
,
2027 struct hlist_node
*hnode
, void *arg
)
2030 struct ldlm_resource
*res
= cfs_hash_object(hs
, hnode
);
2032 return ldlm_resource_foreach(res
, ldlm_iter_helper
, arg
) ==
2036 void ldlm_namespace_foreach(struct ldlm_namespace
*ns
,
2037 ldlm_iterator_t iter
, void *closure
)
2040 struct iter_helper_data helper
= {
2045 cfs_hash_for_each_nolock(ns
->ns_rs_hash
,
2046 ldlm_res_iter_helper
, &helper
);
2049 EXPORT_SYMBOL(ldlm_namespace_foreach
);
2051 /* non-blocking function to manipulate a lock whose cb_data is being put away.
2052 * return 0: find no resource
2053 * > 0: must be LDLM_ITER_STOP/LDLM_ITER_CONTINUE.
2056 int ldlm_resource_iterate(struct ldlm_namespace
*ns
,
2057 const struct ldlm_res_id
*res_id
,
2058 ldlm_iterator_t iter
, void *data
)
2060 struct ldlm_resource
*res
;
2064 CERROR("must pass in namespace\n");
2068 res
= ldlm_resource_get(ns
, NULL
, res_id
, 0, 0);
2072 LDLM_RESOURCE_ADDREF(res
);
2073 rc
= ldlm_resource_foreach(res
, iter
, data
);
2074 LDLM_RESOURCE_DELREF(res
);
2075 ldlm_resource_putref(res
);
2078 EXPORT_SYMBOL(ldlm_resource_iterate
);
2082 static int ldlm_chain_lock_for_replay(struct ldlm_lock
*lock
, void *closure
)
2084 struct list_head
*list
= closure
;
2086 /* we use l_pending_chain here, because it's unused on clients. */
2087 LASSERTF(list_empty(&lock
->l_pending_chain
),
2088 "lock %p next %p prev %p\n",
2089 lock
, &lock
->l_pending_chain
.next
,&lock
->l_pending_chain
.prev
);
2090 /* bug 9573: don't replay locks left after eviction, or
2091 * bug 17614: locks being actively cancelled. Get a reference
2092 * on a lock so that it does not disapear under us (e.g. due to cancel)
2094 if (!(lock
->l_flags
& (LDLM_FL_FAILED
|LDLM_FL_CANCELING
))) {
2095 list_add(&lock
->l_pending_chain
, list
);
2096 LDLM_LOCK_GET(lock
);
2099 return LDLM_ITER_CONTINUE
;
2102 static int replay_lock_interpret(const struct lu_env
*env
,
2103 struct ptlrpc_request
*req
,
2104 struct ldlm_async_args
*aa
, int rc
)
2106 struct ldlm_lock
*lock
;
2107 struct ldlm_reply
*reply
;
2108 struct obd_export
*exp
;
2110 atomic_dec(&req
->rq_import
->imp_replay_inflight
);
2115 reply
= req_capsule_server_get(&req
->rq_pill
, &RMF_DLM_REP
);
2117 GOTO(out
, rc
= -EPROTO
);
2119 lock
= ldlm_handle2lock(&aa
->lock_handle
);
2121 CERROR("received replay ack for unknown local cookie "LPX64
2122 " remote cookie "LPX64
" from server %s id %s\n",
2123 aa
->lock_handle
.cookie
, reply
->lock_handle
.cookie
,
2124 req
->rq_export
->exp_client_uuid
.uuid
,
2125 libcfs_id2str(req
->rq_peer
));
2126 GOTO(out
, rc
= -ESTALE
);
2129 /* Key change rehash lock in per-export hash with new key */
2130 exp
= req
->rq_export
;
2131 if (exp
&& exp
->exp_lock_hash
) {
2132 /* In the function below, .hs_keycmp resolves to
2133 * ldlm_export_lock_keycmp() */
2134 /* coverity[overrun-buffer-val] */
2135 cfs_hash_rehash_key(exp
->exp_lock_hash
,
2136 &lock
->l_remote_handle
,
2137 &reply
->lock_handle
,
2140 lock
->l_remote_handle
= reply
->lock_handle
;
2143 LDLM_DEBUG(lock
, "replayed lock:");
2144 ptlrpc_import_recovery_state_machine(req
->rq_import
);
2145 LDLM_LOCK_PUT(lock
);
2148 ptlrpc_connect_import(req
->rq_import
);
2153 static int replay_one_lock(struct obd_import
*imp
, struct ldlm_lock
*lock
)
2155 struct ptlrpc_request
*req
;
2156 struct ldlm_async_args
*aa
;
2157 struct ldlm_request
*body
;
2160 /* Bug 11974: Do not replay a lock which is actively being canceled */
2161 if (lock
->l_flags
& LDLM_FL_CANCELING
) {
2162 LDLM_DEBUG(lock
, "Not replaying canceled lock:");
2166 /* If this is reply-less callback lock, we cannot replay it, since
2167 * server might have long dropped it, but notification of that event was
2168 * lost by network. (and server granted conflicting lock already) */
2169 if (lock
->l_flags
& LDLM_FL_CANCEL_ON_BLOCK
) {
2170 LDLM_DEBUG(lock
, "Not replaying reply-less lock:");
2171 ldlm_lock_cancel(lock
);
2176 * If granted mode matches the requested mode, this lock is granted.
2178 * If they differ, but we have a granted mode, then we were granted
2179 * one mode and now want another: ergo, converting.
2181 * If we haven't been granted anything and are on a resource list,
2182 * then we're blocked/waiting.
2184 * If we haven't been granted anything and we're NOT on a resource list,
2185 * then we haven't got a reply yet and don't have a known disposition.
2186 * This happens whenever a lock enqueue is the request that triggers
2189 if (lock
->l_granted_mode
== lock
->l_req_mode
)
2190 flags
= LDLM_FL_REPLAY
| LDLM_FL_BLOCK_GRANTED
;
2191 else if (lock
->l_granted_mode
)
2192 flags
= LDLM_FL_REPLAY
| LDLM_FL_BLOCK_CONV
;
2193 else if (!list_empty(&lock
->l_res_link
))
2194 flags
= LDLM_FL_REPLAY
| LDLM_FL_BLOCK_WAIT
;
2196 flags
= LDLM_FL_REPLAY
;
2198 req
= ptlrpc_request_alloc_pack(imp
, &RQF_LDLM_ENQUEUE
,
2199 LUSTRE_DLM_VERSION
, LDLM_ENQUEUE
);
2203 /* We're part of recovery, so don't wait for it. */
2204 req
->rq_send_state
= LUSTRE_IMP_REPLAY_LOCKS
;
2206 body
= req_capsule_client_get(&req
->rq_pill
, &RMF_DLM_REQ
);
2207 ldlm_lock2desc(lock
, &body
->lock_desc
);
2208 body
->lock_flags
= ldlm_flags_to_wire(flags
);
2210 ldlm_lock2handle(lock
, &body
->lock_handle
[0]);
2211 if (lock
->l_lvb_len
> 0)
2212 req_capsule_extend(&req
->rq_pill
, &RQF_LDLM_ENQUEUE_LVB
);
2213 req_capsule_set_size(&req
->rq_pill
, &RMF_DLM_LVB
, RCL_SERVER
,
2215 ptlrpc_request_set_replen(req
);
2216 /* notify the server we've replayed all requests.
2217 * also, we mark the request to be put on a dedicated
2218 * queue to be processed after all request replayes.
2220 lustre_msg_set_flags(req
->rq_reqmsg
, MSG_REQ_REPLAY_DONE
);
2222 LDLM_DEBUG(lock
, "replaying lock:");
2224 atomic_inc(&req
->rq_import
->imp_replay_inflight
);
2225 CLASSERT(sizeof(*aa
) <= sizeof(req
->rq_async_args
));
2226 aa
= ptlrpc_req_async_args(req
);
2227 aa
->lock_handle
= body
->lock_handle
[0];
2228 req
->rq_interpret_reply
= (ptlrpc_interpterer_t
)replay_lock_interpret
;
2229 ptlrpcd_add_req(req
, PDL_POLICY_LOCAL
, -1);
2235 * Cancel as many unused locks as possible before replay. since we are
2236 * in recovery, we can't wait for any outstanding RPCs to send any RPC
2239 * Called only in recovery before replaying locks. there is no need to
2240 * replay locks that are unused. since the clients may hold thousands of
2241 * cached unused locks, dropping the unused locks can greatly reduce the
2242 * load on the servers at recovery time.
2244 static void ldlm_cancel_unused_locks_for_replay(struct ldlm_namespace
*ns
)
2249 CDEBUG(D_DLMTRACE
, "Dropping as many unused locks as possible before"
2250 "replay for namespace %s (%d)\n",
2251 ldlm_ns_name(ns
), ns
->ns_nr_unused
);
2253 /* We don't need to care whether or not LRU resize is enabled
2254 * because the LDLM_CANCEL_NO_WAIT policy doesn't use the
2255 * count parameter */
2256 canceled
= ldlm_cancel_lru_local(ns
, &cancels
, ns
->ns_nr_unused
, 0,
2257 LCF_LOCAL
, LDLM_CANCEL_NO_WAIT
);
2259 CDEBUG(D_DLMTRACE
, "Canceled %d unused locks from namespace %s\n",
2260 canceled
, ldlm_ns_name(ns
));
2263 int ldlm_replay_locks(struct obd_import
*imp
)
2265 struct ldlm_namespace
*ns
= imp
->imp_obd
->obd_namespace
;
2267 struct ldlm_lock
*lock
, *next
;
2270 LASSERT(atomic_read(&imp
->imp_replay_inflight
) == 0);
2272 /* don't replay locks if import failed recovery */
2273 if (imp
->imp_vbr_failed
)
2276 /* ensure this doesn't fall to 0 before all have been queued */
2277 atomic_inc(&imp
->imp_replay_inflight
);
2279 if (ldlm_cancel_unused_locks_before_replay
)
2280 ldlm_cancel_unused_locks_for_replay(ns
);
2282 ldlm_namespace_foreach(ns
, ldlm_chain_lock_for_replay
, &list
);
2284 list_for_each_entry_safe(lock
, next
, &list
, l_pending_chain
) {
2285 list_del_init(&lock
->l_pending_chain
);
2287 LDLM_LOCK_RELEASE(lock
);
2288 continue; /* or try to do the rest? */
2290 rc
= replay_one_lock(imp
, lock
);
2291 LDLM_LOCK_RELEASE(lock
);
2294 atomic_dec(&imp
->imp_replay_inflight
);
2298 EXPORT_SYMBOL(ldlm_replay_locks
);